Reformat code style

app.py

@@ -4,7 +4,8 @@ from gradio_litmodel3d import LitModel3D
 
 import os
 import shutil
-os.environ['SPCONV_ALGO'] = 'native'
+
+os.environ["SPCONV_ALGO"] = "native"
 from typing import *
 import torch
 import numpy as np
@@ -17,15 +18,15 @@ from trellis.utils import render_utils, postprocessing_utils
 
 
 MAX_SEED = np.iinfo(np.int32).max
-TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
+TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "tmp")
 os.makedirs(TMP_DIR, exist_ok=True)
 
 
 def start_session(req: gr.Request):
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     os.makedirs(user_dir, exist_ok=True)
-    
-    
+
+
 def end_session(req: gr.Request):
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     if os.path.exists(user_dir):
@@ -35,7 +36,7 @@ def end_session(req: gr.Request):
 def preprocess_image(image: Image.Image) -> Image.Image:
     """
     Preprocess the input image for 3D generation.
-    
+
     This function is called when a user uploads an image or selects an example.
     It applies background removal and other preprocessing steps necessary for
     optimal 3D model generation.
@@ -53,13 +54,13 @@ def preprocess_image(image: Image.Image) -> Image.Image:
 def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
     """
     Preprocess a list of input images for multi-image 3D generation.
-    
+
     This function is called when users upload multiple images in the gallery.
     It processes each image to prepare them for the multi-image 3D generation pipeline.
-    
+
     Args:
         images (List[Tuple[Image.Image, str]]): The input images from the gallery
-        
+
     Returns:
         List[Image.Image]: The preprocessed images ready for 3D generation
     """
@@ -70,55 +71,55 @@ def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image
 
 def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
     return {
-        'gaussian': {
+        "gaussian": {
             **gs.init_params,
-            '_xyz': gs._xyz.cpu().numpy(),
-            '_features_dc': gs._features_dc.cpu().numpy(),
-            '_scaling': gs._scaling.cpu().numpy(),
-            '_rotation': gs._rotation.cpu().numpy(),
-            '_opacity': gs._opacity.cpu().numpy(),
+            "_xyz": gs._xyz.cpu().numpy(),
+            "_features_dc": gs._features_dc.cpu().numpy(),
+            "_scaling": gs._scaling.cpu().numpy(),
+            "_rotation": gs._rotation.cpu().numpy(),
+            "_opacity": gs._opacity.cpu().numpy(),
         },
-        'mesh': {
-            'vertices': mesh.vertices.cpu().numpy(),
-            'faces': mesh.faces.cpu().numpy(),
+        "mesh": {
+            "vertices": mesh.vertices.cpu().numpy(),
+            "faces": mesh.faces.cpu().numpy(),
         },
     }
-    
-    
+
+
 def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
     gs = Gaussian(
-        aabb=state['gaussian']['aabb'],
-        sh_degree=state['gaussian']['sh_degree'],
-        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],
-        scaling_bias=state['gaussian']['scaling_bias'],
-        opacity_bias=state['gaussian']['opacity_bias'],
-        scaling_activation=state['gaussian']['scaling_activation'],
+        aabb=state["gaussian"]["aabb"],
+        sh_degree=state["gaussian"]["sh_degree"],
+        mininum_kernel_size=state["gaussian"]["mininum_kernel_size"],
+        scaling_bias=state["gaussian"]["scaling_bias"],
+        opacity_bias=state["gaussian"]["opacity_bias"],
+        scaling_activation=state["gaussian"]["scaling_activation"],
     )
-    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda')
-    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda')
-    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda')
-    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda')
-    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda')
-    
+    gs._xyz = torch.tensor(state["gaussian"]["_xyz"], device="cuda")
+    gs._features_dc = torch.tensor(state["gaussian"]["_features_dc"], device="cuda")
+    gs._scaling = torch.tensor(state["gaussian"]["_scaling"], device="cuda")
+    gs._rotation = torch.tensor(state["gaussian"]["_rotation"], device="cuda")
+    gs._opacity = torch.tensor(state["gaussian"]["_opacity"], device="cuda")
+
     mesh = edict(
-        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),
-        faces=torch.tensor(state['mesh']['faces'], device='cuda'),
+        vertices=torch.tensor(state["mesh"]["vertices"], device="cuda"),
+        faces=torch.tensor(state["mesh"]["faces"], device="cuda"),
     )
-    
+
     return gs, mesh
 
 
 def get_seed(randomize_seed: bool, seed: int) -> int:
     """
     Get the random seed for generation.
-    
+
     This function is called by the generate button to determine whether to use
     a random seed or the user-specified seed value.
-    
+
     Args:
         randomize_seed (bool): Whether to generate a random seed
         seed (int): The user-specified seed value
-        
+
     Returns:
         int: The seed to use for generation
     """
@@ -163,7 +164,7 @@ def generate_and_extract_glb(
         str: The path to the extracted GLB file (for download).
     """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
-    
+
     # Generate 3D model
     if not is_multiimage:
         outputs = pipeline.run(
@@ -196,24 +197,28 @@ def generate_and_extract_glb(
             },
             mode=multiimage_algo,
         )
-    
+
     # Render video
-    video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
-    video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
-    video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
-    video_path = os.path.join(user_dir, 'sample.mp4')
+    video = render_utils.render_video(outputs["gaussian"][0], num_frames=120)["color"]
+    video_geo = render_utils.render_video(outputs["mesh"][0], num_frames=120)["normal"]
+    video = [
+        np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))
+    ]
+    video_path = os.path.join(user_dir, "sample.mp4")
     imageio.mimsave(video_path, video, fps=15)
-    
+
     # Extract GLB
-    gs = outputs['gaussian'][0]
-    mesh = outputs['mesh'][0]
-    glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
-    glb_path = os.path.join(user_dir, 'sample.glb')
+    gs = outputs["gaussian"][0]
+    mesh = outputs["mesh"][0]
+    glb = postprocessing_utils.to_glb(
+        gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False
+    )
+    glb_path = os.path.join(user_dir, "sample.glb")
     glb.export(glb_path)
-    
+
     # Pack state for optional Gaussian extraction
     state = pack_state(gs, mesh)
-    
+
     torch.cuda.empty_cache()
     return state, video_path, glb_path, glb_path
 
@@ -222,7 +227,7 @@ def generate_and_extract_glb(
 def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
     """
     Extract a Gaussian splatting file from the generated 3D model.
-    
+
     This function is called when the user clicks "Extract Gaussian" button.
     It converts the 3D model state into a .ply file format containing
     Gaussian splatting data for advanced 3D applications.
@@ -236,7 +241,7 @@ def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
     """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     gs, _ = unpack_state(state)
-    gaussian_path = os.path.join(user_dir, 'sample.ply')
+    gaussian_path = os.path.join(user_dir, "sample.ply")
     gs.save_ply(gaussian_path)
     torch.cuda.empty_cache()
     return gaussian_path, gaussian_path
@@ -250,82 +255,124 @@ def prepare_multi_example() -> List[Image.Image]:
 def split_image(image: Image.Image) -> List[Image.Image]:
     """
     Split a multi-view image into separate view images.
-    
+
     This function is called when users select multi-image examples that contain
     multiple views in a single concatenated image. It automatically splits them
     based on alpha channel boundaries and preprocesses each view.
-    
+
     Args:
         image (Image.Image): A concatenated image containing multiple views
-        
+
     Returns:
         List[Image.Image]: List of individual preprocessed view images
     """
     image = np.array(image)
     alpha = image[..., 3]
-    alpha = np.any(alpha>0, axis=0)
+    alpha = np.any(alpha > 0, axis=0)
     start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()
     end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()
     images = []
     for s, e in zip(start_pos, end_pos):
-        images.append(Image.fromarray(image[:, s:e+1]))
+        images.append(Image.fromarray(image[:, s : e + 1]))
     return [preprocess_image(image) for image in images]
 
 
 with gr.Blocks(delete_cache=(600, 600)) as demo:
-    gr.Markdown("""
+    gr.Markdown(
+        """
     ## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
     * Upload an image and click "Generate & Extract GLB" to create a 3D asset and automatically extract the GLB file.
     * If you want the Gaussian file as well, click "Extract Gaussian" after generation.
     * If the image has alpha channel, it will be used as the mask. Otherwise, we use `rembg` to remove the background.
     
     ✨New: 1) Experimental multi-image support. 2) Gaussian file extraction.
-    """)
-    
+    """
+    )
+
     with gr.Row():
         with gr.Column():
             with gr.Tabs() as input_tabs:
                 with gr.Tab(label="Single Image", id=0) as single_image_input_tab:
-                    image_prompt = gr.Image(label="Image Prompt", format="png", image_mode="RGBA", type="pil", height=300)
+                    image_prompt = gr.Image(
+                        label="Image Prompt",
+                        format="png",
+                        image_mode="RGBA",
+                        type="pil",
+                        height=300,
+                    )
                 with gr.Tab(label="Multiple Images", id=1) as multiimage_input_tab:
-                    multiimage_prompt = gr.Gallery(label="Image Prompt", format="png", type="pil", height=300, columns=3)
-                    gr.Markdown("""
+                    multiimage_prompt = gr.Gallery(
+                        label="Image Prompt",
+                        format="png",
+                        type="pil",
+                        height=300,
+                        columns=3,
+                    )
+                    gr.Markdown(
+                        """
                         Input different views of the object in separate images. 
                         
                         *NOTE: this is an experimental algorithm without training a specialized model. It may not produce the best results for all images, especially those having different poses or inconsistent details.*
-                    """)
-            
+                    """
+                    )
+
             with gr.Accordion(label="Generation Settings", open=False):
                 seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
                 randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                 gr.Markdown("Stage 1: Sparse Structure Generation")
                 with gr.Row():
-                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
-                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
+                    ss_guidance_strength = gr.Slider(
+                        0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1
+                    )
+                    ss_sampling_steps = gr.Slider(
+                        1, 50, label="Sampling Steps", value=12, step=1
+                    )
                 gr.Markdown("Stage 2: Structured Latent Generation")
                 with gr.Row():
-                    slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
-                    slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
-                multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], label="Multi-image Algorithm", value="stochastic")
-            
+                    slat_guidance_strength = gr.Slider(
+                        0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1
+                    )
+                    slat_sampling_steps = gr.Slider(
+                        1, 50, label="Sampling Steps", value=12, step=1
+                    )
+                multiimage_algo = gr.Radio(
+                    ["stochastic", "multidiffusion"],
+                    label="Multi-image Algorithm",
+                    value="stochastic",
+                )
+
             with gr.Accordion(label="GLB Extraction Settings", open=False):
-                mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
-                texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
+                mesh_simplify = gr.Slider(
+                    0.9, 0.98, label="Simplify", value=0.95, step=0.01
+                )
+                texture_size = gr.Slider(
+                    512, 2048, label="Texture Size", value=1024, step=512
+                )
 
             generate_btn = gr.Button("Generate & Extract GLB", variant="primary")
             extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
-            gr.Markdown("""
+            gr.Markdown(
+                """
                         *NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
-                        """)
+                        """
+            )
 
         with gr.Column():
-            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
-            model_output = LitModel3D(label="Extracted GLB/Gaussian", exposure=10.0, height=300)
-            
+            video_output = gr.Video(
+                label="Generated 3D Asset", autoplay=True, loop=True, height=300
+            )
+            model_output = LitModel3D(
+                label="Extracted GLB/Gaussian", exposure=10.0, height=300
+            )
+
             with gr.Row():
-                download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
-                download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)  
-    
+                download_glb = gr.DownloadButton(
+                    label="Download GLB", interactive=False
+                )
+                download_gs = gr.DownloadButton(
+                    label="Download Gaussian", interactive=False
+                )
+
     is_multiimage = gr.State(False)
     output_buf = gr.State()
 
@@ -334,9 +381,9 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
         if os.path.exists("assets/images"):
             examples = gr.Examples(
                 examples=[
-                    f'assets/images/{image}'
+                    f"assets/images/{image}"
                     for image in os.listdir("assets/images")
-                    if image.endswith(('.png', '.jpg', '.jpeg', '.webp'))
+                    if image.endswith((".png", ".jpg", ".jpeg", ".webp"))
                 ],
                 inputs=[image_prompt],
                 fn=preprocess_image,
@@ -346,7 +393,7 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
             )
         else:
             examples = gr.Examples(examples=[], inputs=[image_prompt])
-    
+
     with gr.Row(visible=False) as multiimage_example:
         examples_multi = gr.Examples(
             examples=prepare_multi_example(),
@@ -360,16 +407,20 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
     # Handlers
     demo.load(start_session)
     demo.unload(end_session)
-    
+
     single_image_input_tab.select(
-        lambda: tuple([False, gr.Row.update(visible=True), gr.Row.update(visible=False)]),
-        outputs=[is_multiimage, single_image_example, multiimage_example]
+        lambda: tuple(
+            [False, gr.Row.update(visible=True), gr.Row.update(visible=False)]
+        ),
+        outputs=[is_multiimage, single_image_example, multiimage_example],
     )
     multiimage_input_tab.select(
-        lambda: tuple([True, gr.Row.update(visible=False), gr.Row.update(visible=True)]),
-        outputs=[is_multiimage, single_image_example, multiimage_example]
+        lambda: tuple(
+            [True, gr.Row.update(visible=False), gr.Row.update(visible=True)]
+        ),
+        outputs=[is_multiimage, single_image_example, multiimage_example],
     )
-    
+
     image_prompt.upload(
         preprocess_image,
         inputs=[image_prompt],
@@ -387,7 +438,19 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
         outputs=[seed],
     ).then(
         generate_and_extract_glb,
-        inputs=[image_prompt, multiimage_prompt, is_multiimage, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo, mesh_simplify, texture_size],
+        inputs=[
+            image_prompt,
+            multiimage_prompt,
+            is_multiimage,
+            seed,
+            ss_guidance_strength,
+            ss_sampling_steps,
+            slat_guidance_strength,
+            slat_sampling_steps,
+            multiimage_algo,
+            mesh_simplify,
+            texture_size,
+        ],
         outputs=[output_buf, video_output, model_output, download_glb],
     ).then(
         lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),
@@ -395,10 +458,16 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
     )
 
     video_output.clear(
-        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False), gr.Button(interactive=False)]),
+        lambda: tuple(
+            [
+                gr.Button(interactive=False),
+                gr.Button(interactive=False),
+                gr.Button(interactive=False),
+            ]
+        ),
         outputs=[extract_gs_btn, download_glb, download_gs],
     )
-    
+
     extract_gs_btn.click(
         extract_gaussian,
         inputs=[output_buf],
@@ -412,14 +481,16 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
         lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),
         outputs=[download_glb, download_gs],
     )
-    
+
 
 # Launch the Gradio app
 if __name__ == "__main__":
     pipeline = TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")
     pipeline.cuda()
     try:
-        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))    # Preload rembg
+        pipeline.preprocess_image(
+            Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))
+        )  # Preload rembg
     except:
         pass
     demo.launch()

trellis/models/__init__.py

@@ -1,20 +1,21 @@
 import importlib
 
 __attributes = {
-    'SparseStructureEncoder': 'sparse_structure_vae',
-    'SparseStructureDecoder': 'sparse_structure_vae',
-    'SparseStructureFlowModel': 'sparse_structure_flow',
-    'SLatEncoder': 'structured_latent_vae',
-    'SLatGaussianDecoder': 'structured_latent_vae',
-    'SLatRadianceFieldDecoder': 'structured_latent_vae',
-    'SLatMeshDecoder': 'structured_latent_vae',
-    'SLatFlowModel': 'structured_latent_flow',
+    "SparseStructureEncoder": "sparse_structure_vae",
+    "SparseStructureDecoder": "sparse_structure_vae",
+    "SparseStructureFlowModel": "sparse_structure_flow",
+    "SLatEncoder": "structured_latent_vae",
+    "SLatGaussianDecoder": "structured_latent_vae",
+    "SLatRadianceFieldDecoder": "structured_latent_vae",
+    "SLatMeshDecoder": "structured_latent_vae",
+    "SLatFlowModel": "structured_latent_flow",
 }
 
 __submodules = []
 
 __all__ = list(__attributes.keys()) + __submodules
 
+
 def __getattr__(name):
     if name not in globals():
         if name in __attributes:
@@ -41,6 +42,7 @@ def from_pretrained(path: str, **kwargs):
     import os
     import json
     from safetensors.torch import load_file
+
     is_local = os.path.exists(f"{path}.json") and os.path.exists(f"{path}.safetensors")
 
     if is_local:
@@ -48,23 +50,29 @@ def from_pretrained(path: str, **kwargs):
         model_file = f"{path}.safetensors"
     else:
         from huggingface_hub import hf_hub_download
-        path_parts = path.split('/')
-        repo_id = f'{path_parts[0]}/{path_parts[1]}'
-        model_name = '/'.join(path_parts[2:])
+
+        path_parts = path.split("/")
+        repo_id = f"{path_parts[0]}/{path_parts[1]}"
+        model_name = "/".join(path_parts[2:])
         config_file = hf_hub_download(repo_id, f"{model_name}.json")
         model_file = hf_hub_download(repo_id, f"{model_name}.safetensors")
 
-    with open(config_file, 'r') as f:
+    with open(config_file, "r") as f:
         config = json.load(f)
-    model = __getattr__(config['name'])(**config['args'], **kwargs)
+    model = __getattr__(config["name"])(**config["args"], **kwargs)
     model.load_state_dict(load_file(model_file))
 
     return model
 
 
 # For Pylance
-if __name__ == '__main__':
+if __name__ == "__main__":
     from .sparse_structure_vae import SparseStructureEncoder, SparseStructureDecoder
     from .sparse_structure_flow import SparseStructureFlowModel
-    from .structured_latent_vae import SLatEncoder, SLatGaussianDecoder, SLatRadianceFieldDecoder, SLatMeshDecoder
+    from .structured_latent_vae import (
+        SLatEncoder,
+        SLatGaussianDecoder,
+        SLatRadianceFieldDecoder,
+        SLatMeshDecoder,
+    )
     from .structured_latent_flow import SLatFlowModel

trellis/models/sparse_structure_flow.py

@@ -4,7 +4,10 @@ import torch.nn as nn
 import torch.nn.functional as F
 import numpy as np
 from ..modules.utils import convert_module_to_f16, convert_module_to_f32
-from ..modules.transformer import AbsolutePositionEmbedder, ModulatedTransformerCrossBlock
+from ..modules.transformer import (
+    AbsolutePositionEmbedder,
+    ModulatedTransformerCrossBlock,
+)
 from ..modules.spatial import patchify, unpatchify
 
 
@@ -12,6 +15,7 @@ class TimestepEmbedder(nn.Module):
     """
     Embeds scalar timesteps into vector representations.
     """
+
     def __init__(self, hidden_size, frequency_embedding_size=256):
         super().__init__()
         self.mlp = nn.Sequential(
@@ -38,12 +42,16 @@ class TimestepEmbedder(nn.Module):
         # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
         half = dim // 2
         freqs = torch.exp(
-            -np.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+            -np.log(max_period)
+            * torch.arange(start=0, end=half, dtype=torch.float32)
+            / half
         ).to(device=t.device)
         args = t[:, None].float() * freqs[None]
         embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
         if dim % 2:
-            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+            embedding = torch.cat(
+                [embedding, torch.zeros_like(embedding[:, :1])], dim=-1
+            )
         return embedding
 
     def forward(self, t):
@@ -93,34 +101,41 @@ class SparseStructureFlowModel(nn.Module):
         self.t_embedder = TimestepEmbedder(model_channels)
         if share_mod:
             self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                nn.Linear(model_channels, 6 * model_channels, bias=True)
+                nn.SiLU(), nn.Linear(model_channels, 6 * model_channels, bias=True)
             )
 
         if pe_mode == "ape":
             pos_embedder = AbsolutePositionEmbedder(model_channels, 3)
-            coords = torch.meshgrid(*[torch.arange(res, device=self.device) for res in [resolution // patch_size] * 3], indexing='ij')
+            coords = torch.meshgrid(
+                *[
+                    torch.arange(res, device=self.device)
+                    for res in [resolution // patch_size] * 3
+                ],
+                indexing="ij",
+            )
             coords = torch.stack(coords, dim=-1).reshape(-1, 3)
             pos_emb = pos_embedder(coords)
             self.register_buffer("pos_emb", pos_emb)
 
         self.input_layer = nn.Linear(in_channels * patch_size**3, model_channels)
-            
-        self.blocks = nn.ModuleList([
-            ModulatedTransformerCrossBlock(
-                model_channels,
-                cond_channels,
-                num_heads=self.num_heads,
-                mlp_ratio=self.mlp_ratio,
-                attn_mode='full',
-                use_checkpoint=self.use_checkpoint,
-                use_rope=(pe_mode == "rope"),
-                share_mod=share_mod,
-                qk_rms_norm=self.qk_rms_norm,
-                qk_rms_norm_cross=self.qk_rms_norm_cross,
-            )
-            for _ in range(num_blocks)
-        ])
+
+        self.blocks = nn.ModuleList(
+            [
+                ModulatedTransformerCrossBlock(
+                    model_channels,
+                    cond_channels,
+                    num_heads=self.num_heads,
+                    mlp_ratio=self.mlp_ratio,
+                    attn_mode="full",
+                    use_checkpoint=self.use_checkpoint,
+                    use_rope=(pe_mode == "rope"),
+                    share_mod=share_mod,
+                    qk_rms_norm=self.qk_rms_norm,
+                    qk_rms_norm_cross=self.qk_rms_norm_cross,
+                )
+                for _ in range(num_blocks)
+            ]
+        )
 
         self.out_layer = nn.Linear(model_channels, out_channels * patch_size**3)
 
@@ -154,6 +169,7 @@ class SparseStructureFlowModel(nn.Module):
                 torch.nn.init.xavier_uniform_(module.weight)
                 if module.bias is not None:
                     nn.init.constant_(module.bias, 0)
+
         self.apply(_basic_init)
 
         # Initialize timestep embedding MLP:
@@ -173,9 +189,14 @@ class SparseStructureFlowModel(nn.Module):
         nn.init.constant_(self.out_layer.weight, 0)
         nn.init.constant_(self.out_layer.bias, 0)
 
-    def forward(self, x: torch.Tensor, t: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
-        assert [*x.shape] == [x.shape[0], self.in_channels, *[self.resolution] * 3], \
-                f"Input shape mismatch, got {x.shape}, expected {[x.shape[0], self.in_channels, *[self.resolution] * 3]}"
+    def forward(
+        self, x: torch.Tensor, t: torch.Tensor, cond: torch.Tensor
+    ) -> torch.Tensor:
+        assert [*x.shape] == [
+            x.shape[0],
+            self.in_channels,
+            *[self.resolution] * 3,
+        ], f"Input shape mismatch, got {x.shape}, expected {[x.shape[0], self.in_channels, *[self.resolution] * 3]}"
 
         h = patchify(x, self.patch_size)
         h = h.view(*h.shape[:2], -1).permute(0, 2, 1).contiguous()
@@ -194,7 +215,9 @@ class SparseStructureFlowModel(nn.Module):
         h = F.layer_norm(h, h.shape[-1:])
         h = self.out_layer(h)
 
-        h = h.permute(0, 2, 1).view(h.shape[0], h.shape[2], *[self.resolution // self.patch_size] * 3)
+        h = h.permute(0, 2, 1).view(
+            h.shape[0], h.shape[2], *[self.resolution // self.patch_size] * 3
+        )
         h = unpatchify(h, self.patch_size).contiguous()
 
         return h

trellis/models/sparse_structure_vae.py

@@ -33,9 +33,15 @@ class ResBlock3d(nn.Module):
         self.norm1 = norm_layer(norm_type, channels)
         self.norm2 = norm_layer(norm_type, self.out_channels)
         self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
-        self.conv2 = zero_module(nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1))
-        self.skip_connection = nn.Conv3d(channels, self.out_channels, 1) if channels != self.out_channels else nn.Identity()
-    
+        self.conv2 = zero_module(
+            nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1)
+        )
+        self.skip_connection = (
+            nn.Conv3d(channels, self.out_channels, 1)
+            if channels != self.out_channels
+            else nn.Identity()
+        )
+
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         h = self.norm1(x)
         h = F.silu(h)
@@ -63,7 +69,9 @@ class DownsampleBlock3d(nn.Module):
         if mode == "conv":
             self.conv = nn.Conv3d(in_channels, out_channels, 2, stride=2)
         elif mode == "avgpool":
-            assert in_channels == out_channels, "Pooling mode requires in_channels to be equal to out_channels"
+            assert (
+                in_channels == out_channels
+            ), "Pooling mode requires in_channels to be equal to out_channels"
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         if hasattr(self, "conv"):
@@ -86,9 +94,11 @@ class UpsampleBlock3d(nn.Module):
         self.out_channels = out_channels
 
         if mode == "conv":
-            self.conv = nn.Conv3d(in_channels, out_channels*8, 3, padding=1)
+            self.conv = nn.Conv3d(in_channels, out_channels * 8, 3, padding=1)
         elif mode == "nearest":
-            assert in_channels == out_channels, "Nearest mode requires in_channels to be equal to out_channels"
+            assert (
+                in_channels == out_channels
+            ), "Nearest mode requires in_channels to be equal to out_channels"
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         if hasattr(self, "conv"):
@@ -96,12 +106,12 @@ class UpsampleBlock3d(nn.Module):
             return pixel_shuffle_3d(x, 2)
         else:
             return F.interpolate(x, scale_factor=2, mode="nearest")
-        
+
 
 class SparseStructureEncoder(nn.Module):
     """
     Encoder for Sparse Structure (\mathcal{E}_S in the paper Sec. 3.3).
-    
+
     Args:
         in_channels (int): Channels of the input.
         latent_channels (int): Channels of the latent representation.
@@ -111,6 +121,7 @@ class SparseStructureEncoder(nn.Module):
         norm_type (Literal["group", "layer"]): Type of normalization layer.
         use_fp16 (bool): Whether to use FP16.
     """
+
     def __init__(
         self,
         in_channels: int,
@@ -135,24 +146,21 @@ class SparseStructureEncoder(nn.Module):
 
         self.blocks = nn.ModuleList([])
         for i, ch in enumerate(channels):
-            self.blocks.extend([
-                ResBlock3d(ch, ch)
-                for _ in range(num_res_blocks)
-            ])
+            self.blocks.extend([ResBlock3d(ch, ch) for _ in range(num_res_blocks)])
             if i < len(channels) - 1:
-                self.blocks.append(
-                    DownsampleBlock3d(ch, channels[i+1])
-                )
-        
-        self.middle_block = nn.Sequential(*[
-            ResBlock3d(channels[-1], channels[-1])
-            for _ in range(num_res_blocks_middle)
-        ])
+                self.blocks.append(DownsampleBlock3d(ch, channels[i + 1]))
+
+        self.middle_block = nn.Sequential(
+            *[
+                ResBlock3d(channels[-1], channels[-1])
+                for _ in range(num_res_blocks_middle)
+            ]
+        )
 
         self.out_layer = nn.Sequential(
             norm_layer(norm_type, channels[-1]),
             nn.SiLU(),
-            nn.Conv3d(channels[-1], latent_channels*2, 3, padding=1)
+            nn.Conv3d(channels[-1], latent_channels * 2, 3, padding=1),
         )
 
         if use_fp16:
@@ -183,7 +191,9 @@ class SparseStructureEncoder(nn.Module):
         self.blocks.apply(convert_module_to_f32)
         self.middle_block.apply(convert_module_to_f32)
 
-    def forward(self, x: torch.Tensor, sample_posterior: bool = False, return_raw: bool = False) -> torch.Tensor:
+    def forward(
+        self, x: torch.Tensor, sample_posterior: bool = False, return_raw: bool = False
+    ) -> torch.Tensor:
         h = self.input_layer(x)
         h = h.type(self.dtype)
 
@@ -201,16 +211,16 @@ class SparseStructureEncoder(nn.Module):
             z = mean + std * torch.randn_like(std)
         else:
             z = mean
-            
+
         if return_raw:
             return z, mean, logvar
         return z
-        
+
 
 class SparseStructureDecoder(nn.Module):
     """
     Decoder for Sparse Structure (\mathcal{D}_S in the paper Sec. 3.3).
-    
+
     Args:
         out_channels (int): Channels of the output.
         latent_channels (int): Channels of the latent representation.
@@ -219,7 +229,8 @@ class SparseStructureDecoder(nn.Module):
         num_res_blocks_middle (int): Number of residual blocks in the middle.
         norm_type (Literal["group", "layer"]): Type of normalization layer.
         use_fp16 (bool): Whether to use FP16.
-    """ 
+    """
+
     def __init__(
         self,
         out_channels: int,
@@ -242,26 +253,23 @@ class SparseStructureDecoder(nn.Module):
 
         self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
 
-        self.middle_block = nn.Sequential(*[
-            ResBlock3d(channels[0], channels[0])
-            for _ in range(num_res_blocks_middle)
-        ])
+        self.middle_block = nn.Sequential(
+            *[
+                ResBlock3d(channels[0], channels[0])
+                for _ in range(num_res_blocks_middle)
+            ]
+        )
 
         self.blocks = nn.ModuleList([])
         for i, ch in enumerate(channels):
-            self.blocks.extend([
-                ResBlock3d(ch, ch)
-                for _ in range(num_res_blocks)
-            ])
+            self.blocks.extend([ResBlock3d(ch, ch) for _ in range(num_res_blocks)])
             if i < len(channels) - 1:
-                self.blocks.append(
-                    UpsampleBlock3d(ch, channels[i+1])
-                )
+                self.blocks.append(UpsampleBlock3d(ch, channels[i + 1]))
 
         self.out_layer = nn.Sequential(
             norm_layer(norm_type, channels[-1]),
             nn.SiLU(),
-            nn.Conv3d(channels[-1], out_channels, 3, padding=1)
+            nn.Conv3d(channels[-1], out_channels, 3, padding=1),
         )
 
         if use_fp16:
@@ -273,7 +281,7 @@ class SparseStructureDecoder(nn.Module):
         Return the device of the model.
         """
         return next(self.parameters()).device
-    
+
     def convert_to_fp16(self) -> None:
         """
         Convert the torso of the model to float16.
@@ -291,12 +299,12 @@ class SparseStructureDecoder(nn.Module):
         self.dtype = torch.float32
         self.blocks.apply(convert_module_to_f32)
         self.middle_block.apply(convert_module_to_f32)
-    
+
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         h = self.input_layer(x)
-        
+
         h = h.type(self.dtype)
-                
+
         h = self.middle_block(h)
         for block in self.blocks:
             h = block(h)

trellis/models/structured_latent_flow.py

@@ -26,18 +26,26 @@ class SparseResBlock3d(nn.Module):
         self.out_channels = out_channels or channels
         self.downsample = downsample
         self.upsample = upsample
-        
-        assert not (downsample and upsample), "Cannot downsample and upsample at the same time"
+
+        assert not (
+            downsample and upsample
+        ), "Cannot downsample and upsample at the same time"
 
         self.norm1 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
         self.norm2 = LayerNorm32(self.out_channels, elementwise_affine=False, eps=1e-6)
         self.conv1 = sp.SparseConv3d(channels, self.out_channels, 3)
-        self.conv2 = zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3))
+        self.conv2 = zero_module(
+            sp.SparseConv3d(self.out_channels, self.out_channels, 3)
+        )
         self.emb_layers = nn.Sequential(
             nn.SiLU(),
             nn.Linear(emb_channels, 2 * self.out_channels, bias=True),
         )
-        self.skip_connection = sp.SparseLinear(channels, self.out_channels) if channels != self.out_channels else nn.Identity()
+        self.skip_connection = (
+            sp.SparseLinear(channels, self.out_channels)
+            if channels != self.out_channels
+            else nn.Identity()
+        )
         self.updown = None
         if self.downsample:
             self.updown = sp.SparseDownsample(2)
@@ -63,7 +71,7 @@ class SparseResBlock3d(nn.Module):
         h = h + self.skip_connection(x)
 
         return h
-    
+
 
 class SLatFlowModel(nn.Module):
     def __init__(
@@ -109,14 +117,17 @@ class SLatFlowModel(nn.Module):
         self.qk_rms_norm_cross = qk_rms_norm_cross
         self.dtype = torch.float16 if use_fp16 else torch.float32
 
-        assert int(np.log2(patch_size)) == np.log2(patch_size), "Patch size must be a power of 2"
-        assert np.log2(patch_size) == len(io_block_channels), "Number of IO ResBlocks must match the number of stages"
+        assert int(np.log2(patch_size)) == np.log2(
+            patch_size
+        ), "Patch size must be a power of 2"
+        assert np.log2(patch_size) == len(
+            io_block_channels
+        ), "Number of IO ResBlocks must match the number of stages"
 
         self.t_embedder = TimestepEmbedder(model_channels)
         if share_mod:
             self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                nn.Linear(model_channels, 6 * model_channels, bias=True)
+                nn.SiLU(), nn.Linear(model_channels, 6 * model_channels, bias=True)
             )
 
         if pe_mode == "ape":
@@ -124,15 +135,19 @@ class SLatFlowModel(nn.Module):
 
         self.input_layer = sp.SparseLinear(in_channels, io_block_channels[0])
         self.input_blocks = nn.ModuleList([])
-        for chs, next_chs in zip(io_block_channels, io_block_channels[1:] + [model_channels]):
-            self.input_blocks.extend([
-                SparseResBlock3d(
-                    chs,
-                    model_channels,
-                    out_channels=chs,
-                )
-                for _ in range(num_io_res_blocks-1)
-            ])
+        for chs, next_chs in zip(
+            io_block_channels, io_block_channels[1:] + [model_channels]
+        ):
+            self.input_blocks.extend(
+                [
+                    SparseResBlock3d(
+                        chs,
+                        model_channels,
+                        out_channels=chs,
+                    )
+                    for _ in range(num_io_res_blocks - 1)
+                ]
+            )
             self.input_blocks.append(
                 SparseResBlock3d(
                     chs,
@@ -141,25 +156,30 @@ class SLatFlowModel(nn.Module):
                     downsample=True,
                 )
             )
-            
-        self.blocks = nn.ModuleList([
-            ModulatedSparseTransformerCrossBlock(
-                model_channels,
-                cond_channels,
-                num_heads=self.num_heads,
-                mlp_ratio=self.mlp_ratio,
-                attn_mode='full',
-                use_checkpoint=self.use_checkpoint,
-                use_rope=(pe_mode == "rope"),
-                share_mod=self.share_mod,
-                qk_rms_norm=self.qk_rms_norm,
-                qk_rms_norm_cross=self.qk_rms_norm_cross,
-            )
-            for _ in range(num_blocks)
-        ])
+
+        self.blocks = nn.ModuleList(
+            [
+                ModulatedSparseTransformerCrossBlock(
+                    model_channels,
+                    cond_channels,
+                    num_heads=self.num_heads,
+                    mlp_ratio=self.mlp_ratio,
+                    attn_mode="full",
+                    use_checkpoint=self.use_checkpoint,
+                    use_rope=(pe_mode == "rope"),
+                    share_mod=self.share_mod,
+                    qk_rms_norm=self.qk_rms_norm,
+                    qk_rms_norm_cross=self.qk_rms_norm_cross,
+                )
+                for _ in range(num_blocks)
+            ]
+        )
 
         self.out_blocks = nn.ModuleList([])
-        for chs, prev_chs in zip(reversed(io_block_channels), [model_channels] + list(reversed(io_block_channels[1:]))):
+        for chs, prev_chs in zip(
+            reversed(io_block_channels),
+            [model_channels] + list(reversed(io_block_channels[1:])),
+        ):
             self.out_blocks.append(
                 SparseResBlock3d(
                     prev_chs * 2 if self.use_skip_connection else prev_chs,
@@ -168,14 +188,16 @@ class SLatFlowModel(nn.Module):
                     upsample=True,
                 )
             )
-            self.out_blocks.extend([
-                SparseResBlock3d(
-                    chs * 2 if self.use_skip_connection else chs,
-                    model_channels,
-                    out_channels=chs,
-                )
-                for _ in range(num_io_res_blocks-1)
-            ])
+            self.out_blocks.extend(
+                [
+                    SparseResBlock3d(
+                        chs * 2 if self.use_skip_connection else chs,
+                        model_channels,
+                        out_channels=chs,
+                    )
+                    for _ in range(num_io_res_blocks - 1)
+                ]
+            )
         self.out_layer = sp.SparseLinear(io_block_channels[0], out_channels)
 
         self.initialize_weights()
@@ -212,6 +234,7 @@ class SLatFlowModel(nn.Module):
                 torch.nn.init.xavier_uniform_(module.weight)
                 if module.bias is not None:
                     nn.init.constant_(module.bias, 0)
+
         self.apply(_basic_init)
 
         # Initialize timestep embedding MLP:
@@ -231,7 +254,9 @@ class SLatFlowModel(nn.Module):
         nn.init.constant_(self.out_layer.weight, 0)
         nn.init.constant_(self.out_layer.bias, 0)
 
-    def forward(self, x: sp.SparseTensor, t: torch.Tensor, cond: torch.Tensor) -> sp.SparseTensor:
+    def forward(
+        self, x: sp.SparseTensor, t: torch.Tensor, cond: torch.Tensor
+    ) -> sp.SparseTensor:
         h = self.input_layer(x).type(self.dtype)
         t_emb = self.t_embedder(t)
         if self.share_mod:
@@ -244,7 +269,7 @@ class SLatFlowModel(nn.Module):
         for block in self.input_blocks:
             h = block(h, t_emb)
             skips.append(h.feats)
-        
+
         if self.pe_mode == "ape":
             h = h + self.pos_embedder(h.coords[:, 1:]).type(self.dtype)
         for block in self.blocks:

trellis/models/structured_latent_vae/base.py

@@ -13,15 +13,23 @@ def block_attn_config(self):
     """
     for i in range(self.num_blocks):
         if self.attn_mode == "shift_window":
-            yield "serialized", self.window_size, 0, (16 * (i % 2),) * 3, sp.SerializeMode.Z_ORDER
+            yield "serialized", self.window_size, 0, (
+                16 * (i % 2),
+            ) * 3, sp.SerializeMode.Z_ORDER
         elif self.attn_mode == "shift_sequence":
-            yield "serialized", self.window_size, self.window_size // 2 * (i % 2), (0, 0, 0), sp.SerializeMode.Z_ORDER
+            yield "serialized", self.window_size, self.window_size // 2 * (i % 2), (
+                0,
+                0,
+                0,
+            ), sp.SerializeMode.Z_ORDER
         elif self.attn_mode == "shift_order":
             yield "serialized", self.window_size, 0, (0, 0, 0), sp.SerializeModes[i % 4]
         elif self.attn_mode == "full":
             yield "full", None, None, None, None
         elif self.attn_mode == "swin":
-            yield "windowed", self.window_size, None, self.window_size // 2 * (i % 2), None
+            yield "windowed", self.window_size, None, self.window_size // 2 * (
+                i % 2
+            ), None
 
 
 class SparseTransformerBase(nn.Module):
@@ -29,6 +37,7 @@ class SparseTransformerBase(nn.Module):
     Sparse Transformer without output layers.
     Serve as the base class for encoder and decoder.
     """
+
     def __init__(
         self,
         in_channels: int,
@@ -37,7 +46,9 @@ class SparseTransformerBase(nn.Module):
         num_heads: Optional[int] = None,
         num_head_channels: Optional[int] = 64,
         mlp_ratio: float = 4.0,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "full",
         window_size: Optional[int] = None,
         pe_mode: Literal["ape", "rope"] = "ape",
         use_fp16: bool = False,
@@ -62,22 +73,26 @@ class SparseTransformerBase(nn.Module):
             self.pos_embedder = AbsolutePositionEmbedder(model_channels)
 
         self.input_layer = sp.SparseLinear(in_channels, model_channels)
-        self.blocks = nn.ModuleList([
-            SparseTransformerBlock(
-                model_channels,
-                num_heads=self.num_heads,
-                mlp_ratio=self.mlp_ratio,
-                attn_mode=attn_mode,
-                window_size=window_size,
-                shift_sequence=shift_sequence,
-                shift_window=shift_window,
-                serialize_mode=serialize_mode,
-                use_checkpoint=self.use_checkpoint,
-                use_rope=(pe_mode == "rope"),
-                qk_rms_norm=self.qk_rms_norm,
-            )
-            for attn_mode, window_size, shift_sequence, shift_window, serialize_mode in block_attn_config(self)
-        ])
+        self.blocks = nn.ModuleList(
+            [
+                SparseTransformerBlock(
+                    model_channels,
+                    num_heads=self.num_heads,
+                    mlp_ratio=self.mlp_ratio,
+                    attn_mode=attn_mode,
+                    window_size=window_size,
+                    shift_sequence=shift_sequence,
+                    shift_window=shift_window,
+                    serialize_mode=serialize_mode,
+                    use_checkpoint=self.use_checkpoint,
+                    use_rope=(pe_mode == "rope"),
+                    qk_rms_norm=self.qk_rms_norm,
+                )
+                for attn_mode, window_size, shift_sequence, shift_window, serialize_mode in block_attn_config(
+                    self
+                )
+            ]
+        )
 
     @property
     def device(self) -> torch.device:
@@ -105,6 +120,7 @@ class SparseTransformerBase(nn.Module):
                 torch.nn.init.xavier_uniform_(module.weight)
                 if module.bias is not None:
                     nn.init.constant_(module.bias, 0)
+
         self.apply(_basic_init)
 
     def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:

trellis/models/structured_latent_vae/decoder_gs.py

@@ -18,7 +18,9 @@ class SLatGaussianDecoder(SparseTransformerBase):
         num_heads: Optional[int] = None,
         num_head_channels: Optional[int] = 64,
         mlp_ratio: float = 4,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "swin",
         window_size: int = 8,
         pe_mode: Literal["ape", "rope"] = "ape",
         use_fp16: bool = False,
@@ -57,26 +59,44 @@ class SLatGaussianDecoder(SparseTransformerBase):
         nn.init.constant_(self.out_layer.bias, 0)
 
     def _build_perturbation(self) -> None:
-        perturbation = [hammersley_sequence(3, i, self.rep_config['num_gaussians']) for i in range(self.rep_config['num_gaussians'])]
+        perturbation = [
+            hammersley_sequence(3, i, self.rep_config["num_gaussians"])
+            for i in range(self.rep_config["num_gaussians"])
+        ]
         perturbation = torch.tensor(perturbation).float() * 2 - 1
-        perturbation = perturbation / self.rep_config['voxel_size']
+        perturbation = perturbation / self.rep_config["voxel_size"]
         perturbation = torch.atanh(perturbation).to(self.device)
-        self.register_buffer('offset_perturbation', perturbation)
+        self.register_buffer("offset_perturbation", perturbation)
 
     def _calc_layout(self) -> None:
         self.layout = {
-            '_xyz' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
-            '_features_dc' : {'shape': (self.rep_config['num_gaussians'], 1, 3), 'size': self.rep_config['num_gaussians'] * 3},
-            '_scaling' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
-            '_rotation' : {'shape': (self.rep_config['num_gaussians'], 4), 'size': self.rep_config['num_gaussians'] * 4},
-            '_opacity' : {'shape': (self.rep_config['num_gaussians'], 1), 'size': self.rep_config['num_gaussians']},
+            "_xyz": {
+                "shape": (self.rep_config["num_gaussians"], 3),
+                "size": self.rep_config["num_gaussians"] * 3,
+            },
+            "_features_dc": {
+                "shape": (self.rep_config["num_gaussians"], 1, 3),
+                "size": self.rep_config["num_gaussians"] * 3,
+            },
+            "_scaling": {
+                "shape": (self.rep_config["num_gaussians"], 3),
+                "size": self.rep_config["num_gaussians"] * 3,
+            },
+            "_rotation": {
+                "shape": (self.rep_config["num_gaussians"], 4),
+                "size": self.rep_config["num_gaussians"] * 4,
+            },
+            "_opacity": {
+                "shape": (self.rep_config["num_gaussians"], 1),
+                "size": self.rep_config["num_gaussians"],
+            },
         }
         start = 0
         for k, v in self.layout.items():
-            v['range'] = (start, start + v['size'])
-            start += v['size']
+            v["range"] = (start, start + v["size"])
+            start += v["size"]
         self.out_channels = start
-    
+
     def to_representation(self, x: sp.SparseTensor) -> List[Gaussian]:
         """
         Convert a batch of network outputs to 3D representations.
@@ -92,24 +112,35 @@ class SLatGaussianDecoder(SparseTransformerBase):
             representation = Gaussian(
                 sh_degree=0,
                 aabb=[-0.5, -0.5, -0.5, 1.0, 1.0, 1.0],
-                mininum_kernel_size = self.rep_config['3d_filter_kernel_size'],
-                scaling_bias = self.rep_config['scaling_bias'],
-                opacity_bias = self.rep_config['opacity_bias'],
-                scaling_activation = self.rep_config['scaling_activation']
+                mininum_kernel_size=self.rep_config["3d_filter_kernel_size"],
+                scaling_bias=self.rep_config["scaling_bias"],
+                opacity_bias=self.rep_config["opacity_bias"],
+                scaling_activation=self.rep_config["scaling_activation"],
             )
             xyz = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
             for k, v in self.layout.items():
-                if k == '_xyz':
-                    offset = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape'])
-                    offset = offset * self.rep_config['lr'][k]
-                    if self.rep_config['perturb_offset']:
+                if k == "_xyz":
+                    offset = x.feats[x.layout[i]][
+                        :, v["range"][0] : v["range"][1]
+                    ].reshape(-1, *v["shape"])
+                    offset = offset * self.rep_config["lr"][k]
+                    if self.rep_config["perturb_offset"]:
                         offset = offset + self.offset_perturbation
-                    offset = torch.tanh(offset) / self.resolution * 0.5 * self.rep_config['voxel_size']
+                    offset = (
+                        torch.tanh(offset)
+                        / self.resolution
+                        * 0.5
+                        * self.rep_config["voxel_size"]
+                    )
                     _xyz = xyz.unsqueeze(1) + offset
                     setattr(representation, k, _xyz.flatten(0, 1))
                 else:
-                    feats = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']).flatten(0, 1)
-                    feats = feats * self.rep_config['lr'][k]
+                    feats = (
+                        x.feats[x.layout[i]][:, v["range"][0] : v["range"][1]]
+                        .reshape(-1, *v["shape"])
+                        .flatten(0, 1)
+                    )
+                    feats = feats * self.rep_config["lr"][k]
                     setattr(representation, k, feats)
             ret.append(representation)
         return ret

trellis/models/structured_latent_vae/decoder_mesh.py

@@ -19,12 +19,13 @@ class SparseSubdivideBlock3d(nn.Module):
         out_channels: if specified, the number of output channels.
         num_groups: the number of groups for the group norm.
     """
+
     def __init__(
         self,
         channels: int,
         resolution: int,
         out_channels: Optional[int] = None,
-        num_groups: int = 32
+        num_groups: int = 32,
     ):
         super().__init__()
         self.channels = channels
@@ -33,24 +34,34 @@ class SparseSubdivideBlock3d(nn.Module):
         self.out_channels = out_channels or channels
 
         self.act_layers = nn.Sequential(
-            sp.SparseGroupNorm32(num_groups, channels),
-            sp.SparseSiLU()
+            sp.SparseGroupNorm32(num_groups, channels), sp.SparseSiLU()
         )
-        
+
         self.sub = sp.SparseSubdivide()
-        
+
         self.out_layers = nn.Sequential(
-            sp.SparseConv3d(channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}"),
+            sp.SparseConv3d(
+                channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}"
+            ),
             sp.SparseGroupNorm32(num_groups, self.out_channels),
             sp.SparseSiLU(),
-            zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}")),
+            zero_module(
+                sp.SparseConv3d(
+                    self.out_channels,
+                    self.out_channels,
+                    3,
+                    indice_key=f"res_{self.out_resolution}",
+                )
+            ),
         )
-        
+
         if self.out_channels == channels:
             self.skip_connection = nn.Identity()
         else:
-            self.skip_connection = sp.SparseConv3d(channels, self.out_channels, 1, indice_key=f"res_{self.out_resolution}")
-        
+            self.skip_connection = sp.SparseConv3d(
+                channels, self.out_channels, 1, indice_key=f"res_{self.out_resolution}"
+            )
+
     def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
         """
         Apply the block to a Tensor, conditioned on a timestep embedding.
@@ -78,7 +89,9 @@ class SLatMeshDecoder(SparseTransformerBase):
         num_heads: Optional[int] = None,
         num_head_channels: Optional[int] = 64,
         mlp_ratio: float = 4,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "swin",
         window_size: int = 8,
         pe_mode: Literal["ape", "rope"] = "ape",
         use_fp16: bool = False,
@@ -102,20 +115,24 @@ class SLatMeshDecoder(SparseTransformerBase):
         )
         self.resolution = resolution
         self.rep_config = representation_config
-        self.mesh_extractor = SparseFeatures2Mesh(res=self.resolution*4, use_color=self.rep_config.get('use_color', False))
+        self.mesh_extractor = SparseFeatures2Mesh(
+            res=self.resolution * 4, use_color=self.rep_config.get("use_color", False)
+        )
         self.out_channels = self.mesh_extractor.feats_channels
-        self.upsample = nn.ModuleList([
-            SparseSubdivideBlock3d(
-                channels=model_channels,
-                resolution=resolution,
-                out_channels=model_channels // 4
-            ),
-            SparseSubdivideBlock3d(
-                channels=model_channels // 4,
-                resolution=resolution * 2,
-                out_channels=model_channels // 8
-            )
-        ])
+        self.upsample = nn.ModuleList(
+            [
+                SparseSubdivideBlock3d(
+                    channels=model_channels,
+                    resolution=resolution,
+                    out_channels=model_channels // 4,
+                ),
+                SparseSubdivideBlock3d(
+                    channels=model_channels // 4,
+                    resolution=resolution * 2,
+                    out_channels=model_channels // 8,
+                ),
+            ]
+        )
         self.out_layer = sp.SparseLinear(model_channels // 8, self.out_channels)
 
         self.initialize_weights()
@@ -140,8 +157,8 @@ class SLatMeshDecoder(SparseTransformerBase):
         Convert the torso of the model to float32.
         """
         super().convert_to_fp32()
-        self.upsample.apply(convert_module_to_f32)  
-    
+        self.upsample.apply(convert_module_to_f32)
+
     def to_representation(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
         """
         Convert a batch of network outputs to 3D representations.

trellis/models/structured_latent_vae/decoder_rf.py

@@ -18,7 +18,9 @@ class SLatRadianceFieldDecoder(SparseTransformerBase):
         num_heads: Optional[int] = None,
         num_head_channels: Optional[int] = 64,
         mlp_ratio: float = 4,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "swin",
         window_size: int = 8,
         pe_mode: Literal["ape", "rope"] = "ape",
         use_fp16: bool = False,
@@ -57,16 +59,25 @@ class SLatRadianceFieldDecoder(SparseTransformerBase):
 
     def _calc_layout(self) -> None:
         self.layout = {
-            'trivec': {'shape': (self.rep_config['rank'], 3, self.rep_config['dim']), 'size': self.rep_config['rank'] * 3 * self.rep_config['dim']},
-            'density': {'shape': (self.rep_config['rank'],), 'size': self.rep_config['rank']},
-            'features_dc': {'shape': (self.rep_config['rank'], 1, 3), 'size': self.rep_config['rank'] * 3},
+            "trivec": {
+                "shape": (self.rep_config["rank"], 3, self.rep_config["dim"]),
+                "size": self.rep_config["rank"] * 3 * self.rep_config["dim"],
+            },
+            "density": {
+                "shape": (self.rep_config["rank"],),
+                "size": self.rep_config["rank"],
+            },
+            "features_dc": {
+                "shape": (self.rep_config["rank"], 1, 3),
+                "size": self.rep_config["rank"] * 3,
+            },
         }
         start = 0
         for k, v in self.layout.items():
-            v['range'] = (start, start + v['size'])
-            start += v['size']
-        self.out_channels = start    
-    
+            v["range"] = (start, start + v["size"])
+            start += v["size"]
+        self.out_channels = start
+
     def to_representation(self, x: sp.SparseTensor) -> List[Strivec]:
         """
         Convert a batch of network outputs to 3D representations.
@@ -83,15 +94,28 @@ class SLatRadianceFieldDecoder(SparseTransformerBase):
                 sh_degree=0,
                 resolution=self.resolution,
                 aabb=[-0.5, -0.5, -0.5, 1, 1, 1],
-                rank=self.rep_config['rank'],
-                dim=self.rep_config['dim'],
-                device='cuda',
+                rank=self.rep_config["rank"],
+                dim=self.rep_config["dim"],
+                device="cuda",
             )
             representation.density_shift = 0.0
-            representation.position = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
-            representation.depth = torch.full((representation.position.shape[0], 1), int(np.log2(self.resolution)), dtype=torch.uint8, device='cuda')
+            representation.position = (
+                x.coords[x.layout[i]][:, 1:].float() + 0.5
+            ) / self.resolution
+            representation.depth = torch.full(
+                (representation.position.shape[0], 1),
+                int(np.log2(self.resolution)),
+                dtype=torch.uint8,
+                device="cuda",
+            )
             for k, v in self.layout.items():
-                setattr(representation, k, x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']))
+                setattr(
+                    representation,
+                    k,
+                    x.feats[x.layout[i]][:, v["range"][0] : v["range"][1]].reshape(
+                        -1, *v["shape"]
+                    ),
+                )
             representation.trivec = representation.trivec + 1
             ret.append(representation)
         return ret

trellis/models/structured_latent_vae/encoder.py

@@ -17,7 +17,9 @@ class SLatEncoder(SparseTransformerBase):
         num_heads: Optional[int] = None,
         num_head_channels: Optional[int] = 64,
         mlp_ratio: float = 4,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "swin",
         window_size: int = 8,
         pe_mode: Literal["ape", "rope"] = "ape",
         use_fp16: bool = False,
@@ -56,7 +58,7 @@ class SLatEncoder(SparseTransformerBase):
         h = h.type(x.dtype)
         h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
         h = self.out_layer(h)
-        
+
         # Sample from the posterior distribution
         mean, logvar = h.feats.chunk(2, dim=-1)
         if sample_posterior:
@@ -65,7 +67,7 @@ class SLatEncoder(SparseTransformerBase):
         else:
             z = mean
         z = h.replace(z)
-            
+
         if return_raw:
             return z, mean, logvar
         else:

trellis/modules/attention/__init__.py

@@ -1,32 +1,39 @@
 from typing import *
 
-BACKEND = 'flash_attn' 
+BACKEND = "flash_attn"
 DEBUG = False
 
+
 def __from_env():
     import os
-    
+
     global BACKEND
     global DEBUG
-    
-    env_attn_backend = os.environ.get('ATTN_BACKEND')
-    env_sttn_debug = os.environ.get('ATTN_DEBUG')
-    
-    if env_attn_backend is not None and env_attn_backend in ['xformers', 'flash_attn', 'sdpa', 'naive']:
+
+    env_attn_backend = os.environ.get("ATTN_BACKEND")
+    env_sttn_debug = os.environ.get("ATTN_DEBUG")
+
+    if env_attn_backend is not None and env_attn_backend in [
+        "xformers",
+        "flash_attn",
+        "sdpa",
+        "naive",
+    ]:
         BACKEND = env_attn_backend
     if env_sttn_debug is not None:
-        DEBUG = env_sttn_debug == '1'
+        DEBUG = env_sttn_debug == "1"
 
     print(f"[ATTENTION] Using backend: {BACKEND}")
-        
+
 
 __from_env()
-    
 
-def set_backend(backend: Literal['xformers', 'flash_attn']):
+
+def set_backend(backend: Literal["xformers", "flash_attn"]):
     global BACKEND
     BACKEND = backend
 
+
 def set_debug(debug: bool):
     global DEBUG
     DEBUG = debug

trellis/modules/attention/full_attn.py

@@ -3,20 +3,20 @@ import torch
 import math
 from . import DEBUG, BACKEND
 
-if BACKEND == 'xformers':
+if BACKEND == "xformers":
     import xformers.ops as xops
-elif BACKEND == 'flash_attn':
+elif BACKEND == "flash_attn":
     import flash_attn
-elif BACKEND == 'sdpa':
+elif BACKEND == "sdpa":
     from torch.nn.functional import scaled_dot_product_attention as sdpa
-elif BACKEND == 'naive':
+elif BACKEND == "naive":
     pass
 else:
     raise ValueError(f"Unknown attention backend: {BACKEND}")
 
 
 __all__ = [
-    'scaled_dot_product_attention',
+    "scaled_dot_product_attention",
 ]
 
 
@@ -24,14 +24,14 @@ def _naive_sdpa(q, k, v):
     """
     Naive implementation of scaled dot product attention.
     """
-    q = q.permute(0, 2, 1, 3)   # [N, H, L, C]
-    k = k.permute(0, 2, 1, 3)   # [N, H, L, C]
-    v = v.permute(0, 2, 1, 3)   # [N, H, L, C]
+    q = q.permute(0, 2, 1, 3)  # [N, H, L, C]
+    k = k.permute(0, 2, 1, 3)  # [N, H, L, C]
+    v = v.permute(0, 2, 1, 3)  # [N, H, L, C]
     scale_factor = 1 / math.sqrt(q.size(-1))
     attn_weight = q @ k.transpose(-2, -1) * scale_factor
     attn_weight = torch.softmax(attn_weight, dim=-1)
     out = attn_weight @ v
-    out = out.permute(0, 2, 1, 3)   # [N, L, H, C]
+    out = out.permute(0, 2, 1, 3)  # [N, L, H, C]
     return out
 
 
@@ -45,6 +45,7 @@ def scaled_dot_product_attention(qkv: torch.Tensor) -> torch.Tensor:
     """
     ...
 
+
 @overload
 def scaled_dot_product_attention(q: torch.Tensor, kv: torch.Tensor) -> torch.Tensor:
     """
@@ -56,8 +57,11 @@ def scaled_dot_product_attention(q: torch.Tensor, kv: torch.Tensor) -> torch.Ten
     """
     ...
 
+
 @overload
-def scaled_dot_product_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+def scaled_dot_product_attention(
+    q: torch.Tensor, k: torch.Tensor, v: torch.Tensor
+) -> torch.Tensor:
     """
     Apply scaled dot product attention.
 
@@ -71,64 +75,79 @@ def scaled_dot_product_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tens
     """
     ...
 
+
 def scaled_dot_product_attention(*args, **kwargs):
-    arg_names_dict = {
-        1: ['qkv'],
-        2: ['q', 'kv'],
-        3: ['q', 'k', 'v']
-    }
+    arg_names_dict = {1: ["qkv"], 2: ["q", "kv"], 3: ["q", "k", "v"]}
     num_all_args = len(args) + len(kwargs)
-    assert num_all_args in arg_names_dict, f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
-    for key in arg_names_dict[num_all_args][len(args):]:
+    assert (
+        num_all_args in arg_names_dict
+    ), f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
+    for key in arg_names_dict[num_all_args][len(args) :]:
         assert key in kwargs, f"Missing argument {key}"
 
     if num_all_args == 1:
-        qkv = args[0] if len(args) > 0 else kwargs['qkv']
-        assert len(qkv.shape) == 5 and qkv.shape[2] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, L, 3, H, C]"
+        qkv = args[0] if len(args) > 0 else kwargs["qkv"]
+        assert (
+            len(qkv.shape) == 5 and qkv.shape[2] == 3
+        ), f"Invalid shape for qkv, got {qkv.shape}, expected [N, L, 3, H, C]"
         device = qkv.device
 
     elif num_all_args == 2:
-        q = args[0] if len(args) > 0 else kwargs['q']
-        kv = args[1] if len(args) > 1 else kwargs['kv']
-        assert q.shape[0] == kv.shape[0], f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
-        assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
-        assert len(kv.shape) == 5, f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
+        q = args[0] if len(args) > 0 else kwargs["q"]
+        kv = args[1] if len(args) > 1 else kwargs["kv"]
+        assert (
+            q.shape[0] == kv.shape[0]
+        ), f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
+        assert (
+            len(q.shape) == 4
+        ), f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
+        assert (
+            len(kv.shape) == 5
+        ), f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
         device = q.device
 
     elif num_all_args == 3:
-        q = args[0] if len(args) > 0 else kwargs['q']
-        k = args[1] if len(args) > 1 else kwargs['k']
-        v = args[2] if len(args) > 2 else kwargs['v']
-        assert q.shape[0] == k.shape[0] == v.shape[0], f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
-        assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
-        assert len(k.shape) == 4, f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
-        assert len(v.shape) == 4, f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
-        device = q.device    
-
-    if BACKEND == 'xformers':
+        q = args[0] if len(args) > 0 else kwargs["q"]
+        k = args[1] if len(args) > 1 else kwargs["k"]
+        v = args[2] if len(args) > 2 else kwargs["v"]
+        assert (
+            q.shape[0] == k.shape[0] == v.shape[0]
+        ), f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
+        assert (
+            len(q.shape) == 4
+        ), f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
+        assert (
+            len(k.shape) == 4
+        ), f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
+        assert (
+            len(v.shape) == 4
+        ), f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
+        device = q.device
+
+    if BACKEND == "xformers":
         if num_all_args == 1:
             q, k, v = qkv.unbind(dim=2)
         elif num_all_args == 2:
             k, v = kv.unbind(dim=2)
         out = xops.memory_efficient_attention(q, k, v)
-    elif BACKEND == 'flash_attn':
+    elif BACKEND == "flash_attn":
         if num_all_args == 1:
             out = flash_attn.flash_attn_qkvpacked_func(qkv)
         elif num_all_args == 2:
             out = flash_attn.flash_attn_kvpacked_func(q, kv)
         elif num_all_args == 3:
             out = flash_attn.flash_attn_func(q, k, v)
-    elif BACKEND == 'sdpa':
+    elif BACKEND == "sdpa":
         if num_all_args == 1:
             q, k, v = qkv.unbind(dim=2)
         elif num_all_args == 2:
             k, v = kv.unbind(dim=2)
-        q = q.permute(0, 2, 1, 3)   # [N, H, L, C]
-        k = k.permute(0, 2, 1, 3)   # [N, H, L, C]
-        v = v.permute(0, 2, 1, 3)   # [N, H, L, C]
-        out = sdpa(q, k, v)         # [N, H, L, C]
-        out = out.permute(0, 2, 1, 3)   # [N, L, H, C]
-    elif BACKEND == 'naive':
+        q = q.permute(0, 2, 1, 3)  # [N, H, L, C]
+        k = k.permute(0, 2, 1, 3)  # [N, H, L, C]
+        v = v.permute(0, 2, 1, 3)  # [N, H, L, C]
+        out = sdpa(q, k, v)  # [N, H, L, C]
+        out = out.permute(0, 2, 1, 3)  # [N, L, H, C]
+    elif BACKEND == "naive":
         if num_all_args == 1:
             q, k, v = qkv.unbind(dim=2)
         elif num_all_args == 2:
@@ -136,5 +155,5 @@ def scaled_dot_product_attention(*args, **kwargs):
         out = _naive_sdpa(q, k, v)
     else:
         raise ValueError(f"Unknown attention module: {BACKEND}")
-    
+
     return out

trellis/modules/attention/modules.py

@@ -8,11 +8,11 @@ from .full_attn import scaled_dot_product_attention
 class MultiHeadRMSNorm(nn.Module):
     def __init__(self, dim: int, heads: int):
         super().__init__()
-        self.scale = dim ** 0.5
+        self.scale = dim**0.5
         self.gamma = nn.Parameter(torch.ones(heads, dim))
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
-        return (F.normalize(x.float(), dim = -1) * self.gamma * self.scale).to(x.dtype)
+        return (F.normalize(x.float(), dim=-1) * self.gamma * self.scale).to(x.dtype)
 
 
 class RotaryPositionEmbedder(nn.Module):
@@ -23,21 +23,25 @@ class RotaryPositionEmbedder(nn.Module):
         self.in_channels = in_channels
         self.freq_dim = hidden_size // in_channels // 2
         self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
-        self.freqs = 1.0 / (10000 ** self.freqs)
-        
+        self.freqs = 1.0 / (10000**self.freqs)
+
     def _get_phases(self, indices: torch.Tensor) -> torch.Tensor:
         self.freqs = self.freqs.to(indices.device)
         phases = torch.outer(indices, self.freqs)
         phases = torch.polar(torch.ones_like(phases), phases)
         return phases
-        
+
     def _rotary_embedding(self, x: torch.Tensor, phases: torch.Tensor) -> torch.Tensor:
         x_complex = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
         x_rotated = x_complex * phases
-        x_embed = torch.view_as_real(x_rotated).reshape(*x_rotated.shape[:-1], -1).to(x.dtype)
+        x_embed = (
+            torch.view_as_real(x_rotated).reshape(*x_rotated.shape[:-1], -1).to(x.dtype)
+        )
         return x_embed
-        
-    def forward(self, q: torch.Tensor, k: torch.Tensor, indices: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
+
+    def forward(
+        self, q: torch.Tensor, k: torch.Tensor, indices: Optional[torch.Tensor] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
         """
         Args:
             q (sp.SparseTensor): [..., N, D] tensor of queries
@@ -48,24 +52,38 @@ class RotaryPositionEmbedder(nn.Module):
             indices = torch.arange(q.shape[-2], device=q.device)
             if len(q.shape) > 2:
                 indices = indices.unsqueeze(0).expand(q.shape[:-2] + (-1,))
-        
+
         phases = self._get_phases(indices.reshape(-1)).reshape(*indices.shape[:-1], -1)
         if phases.shape[1] < self.hidden_size // 2:
-            phases = torch.cat([phases, torch.polar(
-                torch.ones(*phases.shape[:-1], self.hidden_size // 2 - phases.shape[1], device=phases.device),
-                torch.zeros(*phases.shape[:-1], self.hidden_size // 2 - phases.shape[1], device=phases.device)
-            )], dim=-1)
+            phases = torch.cat(
+                [
+                    phases,
+                    torch.polar(
+                        torch.ones(
+                            *phases.shape[:-1],
+                            self.hidden_size // 2 - phases.shape[1],
+                            device=phases.device,
+                        ),
+                        torch.zeros(
+                            *phases.shape[:-1],
+                            self.hidden_size // 2 - phases.shape[1],
+                            device=phases.device,
+                        ),
+                    ),
+                ],
+                dim=-1,
+            )
         q_embed = self._rotary_embedding(q, phases)
         k_embed = self._rotary_embedding(k, phases)
         return q_embed, k_embed
-    
+
 
 class MultiHeadAttention(nn.Module):
     def __init__(
         self,
         channels: int,
         num_heads: int,
-        ctx_channels: Optional[int]=None,
+        ctx_channels: Optional[int] = None,
         type: Literal["self", "cross"] = "self",
         attn_mode: Literal["full", "windowed"] = "full",
         window_size: Optional[int] = None,
@@ -78,11 +96,13 @@ class MultiHeadAttention(nn.Module):
         assert channels % num_heads == 0
         assert type in ["self", "cross"], f"Invalid attention type: {type}"
         assert attn_mode in ["full", "windowed"], f"Invalid attention mode: {attn_mode}"
-        assert type == "self" or attn_mode == "full", "Cross-attention only supports full attention"
-        
+        assert (
+            type == "self" or attn_mode == "full"
+        ), "Cross-attention only supports full attention"
+
         if attn_mode == "windowed":
             raise NotImplementedError("Windowed attention is not yet implemented")
-        
+
         self.channels = channels
         self.head_dim = channels // num_heads
         self.ctx_channels = ctx_channels if ctx_channels is not None else channels
@@ -99,17 +119,22 @@ class MultiHeadAttention(nn.Module):
         else:
             self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
             self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
-            
+
         if self.qk_rms_norm:
             self.q_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
             self.k_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
-            
+
         self.to_out = nn.Linear(channels, channels)
 
         if use_rope:
             self.rope = RotaryPositionEmbedder(channels)
-    
-    def forward(self, x: torch.Tensor, context: Optional[torch.Tensor] = None, indices: Optional[torch.Tensor] = None) -> torch.Tensor:
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        indices: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
         B, L, C = x.shape
         if self._type == "self":
             qkv = self.to_qkv(x)

trellis/modules/norm.py

@@ -5,21 +5,21 @@ import torch.nn as nn
 class LayerNorm32(nn.LayerNorm):
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         return super().forward(x.float()).type(x.dtype)
-    
+
 
 class GroupNorm32(nn.GroupNorm):
     """
     A GroupNorm layer that converts to float32 before the forward pass.
     """
+
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         return super().forward(x.float()).type(x.dtype)
-    
-    
+
+
 class ChannelLayerNorm32(LayerNorm32):
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         DIM = x.dim()
         x = x.permute(0, *range(2, DIM), 1).contiguous()
         x = super().forward(x)
-        x = x.permute(0, DIM-1, *range(1, DIM-1)).contiguous()
+        x = x.permute(0, DIM - 1, *range(1, DIM - 1)).contiguous()
         return x
-    

trellis/modules/sparse/__init__.py

@@ -1,81 +1,88 @@
 from typing import *
 
-BACKEND = 'spconv' 
+BACKEND = "spconv"
 DEBUG = False
-ATTN = 'flash_attn'
+ATTN = "flash_attn"
+
 
 def __from_env():
     import os
-    
+
     global BACKEND
     global DEBUG
     global ATTN
-    
-    env_sparse_backend = os.environ.get('SPARSE_BACKEND')
-    env_sparse_debug = os.environ.get('SPARSE_DEBUG')
-    env_sparse_attn = os.environ.get('SPARSE_ATTN_BACKEND')
+
+    env_sparse_backend = os.environ.get("SPARSE_BACKEND")
+    env_sparse_debug = os.environ.get("SPARSE_DEBUG")
+    env_sparse_attn = os.environ.get("SPARSE_ATTN_BACKEND")
     if env_sparse_attn is None:
-        env_sparse_attn = os.environ.get('ATTN_BACKEND')
+        env_sparse_attn = os.environ.get("ATTN_BACKEND")
 
-    if env_sparse_backend is not None and env_sparse_backend in ['spconv', 'torchsparse']:
+    if env_sparse_backend is not None and env_sparse_backend in [
+        "spconv",
+        "torchsparse",
+    ]:
         BACKEND = env_sparse_backend
     if env_sparse_debug is not None:
-        DEBUG = env_sparse_debug == '1'
-    if env_sparse_attn is not None and env_sparse_attn in ['xformers', 'flash_attn']:
+        DEBUG = env_sparse_debug == "1"
+    if env_sparse_attn is not None and env_sparse_attn in ["xformers", "flash_attn"]:
         ATTN = env_sparse_attn
-        
+
     print(f"[SPARSE] Backend: {BACKEND}, Attention: {ATTN}")
-        
+
 
 __from_env()
-    
 
-def set_backend(backend: Literal['spconv', 'torchsparse']):
+
+def set_backend(backend: Literal["spconv", "torchsparse"]):
     global BACKEND
     BACKEND = backend
 
+
 def set_debug(debug: bool):
     global DEBUG
     DEBUG = debug
 
-def set_attn(attn: Literal['xformers', 'flash_attn']):
+
+def set_attn(attn: Literal["xformers", "flash_attn"]):
     global ATTN
     ATTN = attn
-    
-    
+
+
 import importlib
 
 __attributes = {
-    'SparseTensor': 'basic',
-    'sparse_batch_broadcast': 'basic',
-    'sparse_batch_op': 'basic',
-    'sparse_cat': 'basic',
-    'sparse_unbind': 'basic',
-    'SparseGroupNorm': 'norm',
-    'SparseLayerNorm': 'norm',
-    'SparseGroupNorm32': 'norm',
-    'SparseLayerNorm32': 'norm',
-    'SparseReLU': 'nonlinearity',
-    'SparseSiLU': 'nonlinearity',
-    'SparseGELU': 'nonlinearity',
-    'SparseActivation': 'nonlinearity',
-    'SparseLinear': 'linear',
-    'sparse_scaled_dot_product_attention': 'attention',
-    'SerializeMode': 'attention',
-    'sparse_serialized_scaled_dot_product_self_attention': 'attention',
-    'sparse_windowed_scaled_dot_product_self_attention': 'attention',
-    'SparseMultiHeadAttention': 'attention',
-    'SparseConv3d': 'conv',
-    'SparseInverseConv3d': 'conv',
-    'SparseDownsample': 'spatial',
-    'SparseUpsample': 'spatial',
-    'SparseSubdivide' : 'spatial'
+    "SparseTensor": "basic",
+    "sparse_batch_broadcast": "basic",
+    "sparse_batch_op": "basic",
+    "sparse_cat": "basic",
+    "sparse_unbind": "basic",
+    "SparseGroupNorm": "norm",
+    "SparseLayerNorm": "norm",
+    "SparseGroupNorm32": "norm",
+    "SparseLayerNorm32": "norm",
+    "SparseReLU": "nonlinearity",
+    "SparseSiLU": "nonlinearity",
+    "SparseGELU": "nonlinearity",
+    "SparseActivation": "nonlinearity",
+    "SparseLinear": "linear",
+    "sparse_scaled_dot_product_attention": "attention",
+    "SerializeMode": "attention",
+    "sparse_serialized_scaled_dot_product_self_attention": "attention",
+    "sparse_windowed_scaled_dot_product_self_attention": "attention",
+    "SparseMultiHeadAttention": "attention",
+    "SparseConv3d": "conv",
+    "SparseInverseConv3d": "conv",
+    "SparseDownsample": "spatial",
+    "SparseUpsample": "spatial",
+    "SparseSubdivide": "spatial",
 }
 
-__submodules = ['transformer']
+__submodules = ["transformer"]
 
 __all__ = list(__attributes.keys()) + __submodules
 
+
 def __getattr__(name):
     if name not in globals():
         if name in __attributes:
@@ -91,7 +98,7 @@ def __getattr__(name):
 
 
 # For Pylance
-if __name__ == '__main__':
+if __name__ == "__main__":
     from .basic import *
     from .norm import *
     from .nonlinearity import *

trellis/modules/sparse/attention/full_attn.py

@@ -3,16 +3,16 @@ import torch
 from .. import SparseTensor
 from .. import DEBUG, ATTN
 
-if ATTN == 'xformers':
+if ATTN == "xformers":
     import xformers.ops as xops
-elif ATTN == 'flash_attn':
+elif ATTN == "flash_attn":
     import flash_attn
 else:
     raise ValueError(f"Unknown attention module: {ATTN}")
 
 
 __all__ = [
-    'sparse_scaled_dot_product_attention',
+    "sparse_scaled_dot_product_attention",
 ]
 
 
@@ -26,8 +26,11 @@ def sparse_scaled_dot_product_attention(qkv: SparseTensor) -> SparseTensor:
     """
     ...
 
+
 @overload
-def sparse_scaled_dot_product_attention(q: SparseTensor, kv: Union[SparseTensor, torch.Tensor]) -> SparseTensor:
+def sparse_scaled_dot_product_attention(
+    q: SparseTensor, kv: Union[SparseTensor, torch.Tensor]
+) -> SparseTensor:
     """
     Apply scaled dot product attention to a sparse tensor.
 
@@ -37,8 +40,11 @@ def sparse_scaled_dot_product_attention(q: SparseTensor, kv: Union[SparseTensor,
     """
     ...
 
+
 @overload
-def sparse_scaled_dot_product_attention(q: torch.Tensor, kv: SparseTensor) -> torch.Tensor:
+def sparse_scaled_dot_product_attention(
+    q: torch.Tensor, kv: SparseTensor
+) -> torch.Tensor:
     """
     Apply scaled dot product attention to a sparse tensor.
 
@@ -48,8 +54,11 @@ def sparse_scaled_dot_product_attention(q: torch.Tensor, kv: SparseTensor) -> to
     """
     ...
 
+
 @overload
-def sparse_scaled_dot_product_attention(q: SparseTensor, k: SparseTensor, v: SparseTensor) -> SparseTensor:
+def sparse_scaled_dot_product_attention(
+    q: SparseTensor, k: SparseTensor, v: SparseTensor
+) -> SparseTensor:
     """
     Apply scaled dot product attention to a sparse tensor.
 
@@ -63,8 +72,11 @@ def sparse_scaled_dot_product_attention(q: SparseTensor, k: SparseTensor, v: Spa
     """
     ...
 
+
 @overload
-def sparse_scaled_dot_product_attention(q: SparseTensor, k: torch.Tensor, v: torch.Tensor) -> SparseTensor:
+def sparse_scaled_dot_product_attention(
+    q: SparseTensor, k: torch.Tensor, v: torch.Tensor
+) -> SparseTensor:
     """
     Apply scaled dot product attention to a sparse tensor.
 
@@ -75,8 +87,11 @@ def sparse_scaled_dot_product_attention(q: SparseTensor, k: torch.Tensor, v: tor
     """
     ...
 
+
 @overload
-def sparse_scaled_dot_product_attention(q: torch.Tensor, k: SparseTensor, v: SparseTensor) -> torch.Tensor:
+def sparse_scaled_dot_product_attention(
+    q: torch.Tensor, k: SparseTensor, v: SparseTensor
+) -> torch.Tensor:
     """
     Apply scaled dot product attention to a sparse tensor.
 
@@ -87,106 +102,158 @@ def sparse_scaled_dot_product_attention(q: torch.Tensor, k: SparseTensor, v: Spa
     """
     ...
 
+
 def sparse_scaled_dot_product_attention(*args, **kwargs):
-    arg_names_dict = {
-        1: ['qkv'],
-        2: ['q', 'kv'],
-        3: ['q', 'k', 'v']
-    }
+    arg_names_dict = {1: ["qkv"], 2: ["q", "kv"], 3: ["q", "k", "v"]}
     num_all_args = len(args) + len(kwargs)
-    assert num_all_args in arg_names_dict, f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
-    for key in arg_names_dict[num_all_args][len(args):]:
+    assert (
+        num_all_args in arg_names_dict
+    ), f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
+    for key in arg_names_dict[num_all_args][len(args) :]:
         assert key in kwargs, f"Missing argument {key}"
 
     if num_all_args == 1:
-        qkv = args[0] if len(args) > 0 else kwargs['qkv']
-        assert isinstance(qkv, SparseTensor), f"qkv must be a SparseTensor, got {type(qkv)}"
-        assert len(qkv.shape) == 4 and qkv.shape[1] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
+        qkv = args[0] if len(args) > 0 else kwargs["qkv"]
+        assert isinstance(
+            qkv, SparseTensor
+        ), f"qkv must be a SparseTensor, got {type(qkv)}"
+        assert (
+            len(qkv.shape) == 4 and qkv.shape[1] == 3
+        ), f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
         device = qkv.device
 
         s = qkv
-        q_seqlen = [qkv.layout[i].stop - qkv.layout[i].start for i in range(qkv.shape[0])]
+        q_seqlen = [
+            qkv.layout[i].stop - qkv.layout[i].start for i in range(qkv.shape[0])
+        ]
         kv_seqlen = q_seqlen
-        qkv = qkv.feats     # [T, 3, H, C]
+        qkv = qkv.feats  # [T, 3, H, C]
 
     elif num_all_args == 2:
-        q = args[0] if len(args) > 0 else kwargs['q']
-        kv = args[1] if len(args) > 1 else kwargs['kv']
-        assert isinstance(q, SparseTensor) and isinstance(kv, (SparseTensor, torch.Tensor)) or \
-               isinstance(q, torch.Tensor) and isinstance(kv, SparseTensor), \
-               f"Invalid types, got {type(q)} and {type(kv)}"
-        assert q.shape[0] == kv.shape[0], f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
+        q = args[0] if len(args) > 0 else kwargs["q"]
+        kv = args[1] if len(args) > 1 else kwargs["kv"]
+        assert (
+            isinstance(q, SparseTensor)
+            and isinstance(kv, (SparseTensor, torch.Tensor))
+            or isinstance(q, torch.Tensor)
+            and isinstance(kv, SparseTensor)
+        ), f"Invalid types, got {type(q)} and {type(kv)}"
+        assert (
+            q.shape[0] == kv.shape[0]
+        ), f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
         device = q.device
 
         if isinstance(q, SparseTensor):
-            assert len(q.shape) == 3, f"Invalid shape for q, got {q.shape}, expected [N, *, H, C]"
+            assert (
+                len(q.shape) == 3
+            ), f"Invalid shape for q, got {q.shape}, expected [N, *, H, C]"
             s = q
             q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
-            q = q.feats     # [T_Q, H, C]
+            q = q.feats  # [T_Q, H, C]
         else:
-            assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
+            assert (
+                len(q.shape) == 4
+            ), f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
             s = None
             N, L, H, C = q.shape
             q_seqlen = [L] * N
-            q = q.reshape(N * L, H, C)   # [T_Q, H, C]
+            q = q.reshape(N * L, H, C)  # [T_Q, H, C]
 
         if isinstance(kv, SparseTensor):
-            assert len(kv.shape) == 4 and kv.shape[1] == 2, f"Invalid shape for kv, got {kv.shape}, expected [N, *, 2, H, C]"
-            kv_seqlen = [kv.layout[i].stop - kv.layout[i].start for i in range(kv.shape[0])]
-            kv = kv.feats     # [T_KV, 2, H, C]
+            assert (
+                len(kv.shape) == 4 and kv.shape[1] == 2
+            ), f"Invalid shape for kv, got {kv.shape}, expected [N, *, 2, H, C]"
+            kv_seqlen = [
+                kv.layout[i].stop - kv.layout[i].start for i in range(kv.shape[0])
+            ]
+            kv = kv.feats  # [T_KV, 2, H, C]
         else:
-            assert len(kv.shape) == 5, f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
+            assert (
+                len(kv.shape) == 5
+            ), f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
             N, L, _, H, C = kv.shape
             kv_seqlen = [L] * N
-            kv = kv.reshape(N * L, 2, H, C)   # [T_KV, 2, H, C]
+            kv = kv.reshape(N * L, 2, H, C)  # [T_KV, 2, H, C]
 
     elif num_all_args == 3:
-        q = args[0] if len(args) > 0 else kwargs['q']
-        k = args[1] if len(args) > 1 else kwargs['k']
-        v = args[2] if len(args) > 2 else kwargs['v']
-        assert isinstance(q, SparseTensor) and isinstance(k, (SparseTensor, torch.Tensor)) and type(k) == type(v) or \
-               isinstance(q, torch.Tensor) and isinstance(k, SparseTensor) and isinstance(v, SparseTensor), \
-               f"Invalid types, got {type(q)}, {type(k)}, and {type(v)}"
-        assert q.shape[0] == k.shape[0] == v.shape[0], f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
+        q = args[0] if len(args) > 0 else kwargs["q"]
+        k = args[1] if len(args) > 1 else kwargs["k"]
+        v = args[2] if len(args) > 2 else kwargs["v"]
+        assert (
+            isinstance(q, SparseTensor)
+            and isinstance(k, (SparseTensor, torch.Tensor))
+            and type(k) == type(v)
+            or isinstance(q, torch.Tensor)
+            and isinstance(k, SparseTensor)
+            and isinstance(v, SparseTensor)
+        ), f"Invalid types, got {type(q)}, {type(k)}, and {type(v)}"
+        assert (
+            q.shape[0] == k.shape[0] == v.shape[0]
+        ), f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
         device = q.device
 
         if isinstance(q, SparseTensor):
-            assert len(q.shape) == 3, f"Invalid shape for q, got {q.shape}, expected [N, *, H, Ci]"
+            assert (
+                len(q.shape) == 3
+            ), f"Invalid shape for q, got {q.shape}, expected [N, *, H, Ci]"
             s = q
             q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
-            q = q.feats     # [T_Q, H, Ci]
+            q = q.feats  # [T_Q, H, Ci]
         else:
-            assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
+            assert (
+                len(q.shape) == 4
+            ), f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
             s = None
             N, L, H, CI = q.shape
             q_seqlen = [L] * N
             q = q.reshape(N * L, H, CI)  # [T_Q, H, Ci]
 
         if isinstance(k, SparseTensor):
-            assert len(k.shape) == 3, f"Invalid shape for k, got {k.shape}, expected [N, *, H, Ci]"
-            assert len(v.shape) == 3, f"Invalid shape for v, got {v.shape}, expected [N, *, H, Co]"
-            kv_seqlen = [k.layout[i].stop - k.layout[i].start for i in range(k.shape[0])]
-            k = k.feats     # [T_KV, H, Ci]
-            v = v.feats     # [T_KV, H, Co]
+            assert (
+                len(k.shape) == 3
+            ), f"Invalid shape for k, got {k.shape}, expected [N, *, H, Ci]"
+            assert (
+                len(v.shape) == 3
+            ), f"Invalid shape for v, got {v.shape}, expected [N, *, H, Co]"
+            kv_seqlen = [
+                k.layout[i].stop - k.layout[i].start for i in range(k.shape[0])
+            ]
+            k = k.feats  # [T_KV, H, Ci]
+            v = v.feats  # [T_KV, H, Co]
         else:
-            assert len(k.shape) == 4, f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
-            assert len(v.shape) == 4, f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
+            assert (
+                len(k.shape) == 4
+            ), f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
+            assert (
+                len(v.shape) == 4
+            ), f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
             N, L, H, CI, CO = *k.shape, v.shape[-1]
             kv_seqlen = [L] * N
-            k = k.reshape(N * L, H, CI)     # [T_KV, H, Ci]
-            v = v.reshape(N * L, H, CO)     # [T_KV, H, Co]
+            k = k.reshape(N * L, H, CI)  # [T_KV, H, Ci]
+            v = v.reshape(N * L, H, CO)  # [T_KV, H, Co]
 
     if DEBUG:
         if s is not None:
             for i in range(s.shape[0]):
-                assert (s.coords[s.layout[i]] == i).all(), f"SparseScaledDotProductSelfAttention: batch index mismatch"
+                assert (
+                    s.coords[s.layout[i]] == i
+                ).all(), f"SparseScaledDotProductSelfAttention: batch index mismatch"
         if num_all_args in [2, 3]:
-            assert q.shape[:2] == [1, sum(q_seqlen)], f"SparseScaledDotProductSelfAttention: q shape mismatch"
+            assert q.shape[:2] == [
+                1,
+                sum(q_seqlen),
+            ], f"SparseScaledDotProductSelfAttention: q shape mismatch"
         if num_all_args == 3:
-            assert k.shape[:2] == [1, sum(kv_seqlen)], f"SparseScaledDotProductSelfAttention: k shape mismatch"
-            assert v.shape[:2] == [1, sum(kv_seqlen)], f"SparseScaledDotProductSelfAttention: v shape mismatch"
+            assert k.shape[:2] == [
+                1,
+                sum(kv_seqlen),
+            ], f"SparseScaledDotProductSelfAttention: k shape mismatch"
+            assert v.shape[:2] == [
+                1,
+                sum(kv_seqlen),
+            ], f"SparseScaledDotProductSelfAttention: v shape mismatch"
 
-    if ATTN == 'xformers':
+    if ATTN == "xformers":
         if num_all_args == 1:
             q, k, v = qkv.unbind(dim=1)
         elif num_all_args == 2:
@@ -196,19 +263,35 @@ def sparse_scaled_dot_product_attention(*args, **kwargs):
         v = v.unsqueeze(0)
         mask = xops.fmha.BlockDiagonalMask.from_seqlens(q_seqlen, kv_seqlen)
         out = xops.memory_efficient_attention(q, k, v, mask)[0]
-    elif ATTN == 'flash_attn':
-        cu_seqlens_q = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(q_seqlen), dim=0)]).int().to(device)
+    elif ATTN == "flash_attn":
+        cu_seqlens_q = (
+            torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(q_seqlen), dim=0)])
+            .int()
+            .to(device)
+        )
         if num_all_args in [2, 3]:
-            cu_seqlens_kv = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(kv_seqlen), dim=0)]).int().to(device)
+            cu_seqlens_kv = (
+                torch.cat(
+                    [torch.tensor([0]), torch.cumsum(torch.tensor(kv_seqlen), dim=0)]
+                )
+                .int()
+                .to(device)
+            )
         if num_all_args == 1:
-            out = flash_attn.flash_attn_varlen_qkvpacked_func(qkv, cu_seqlens_q, max(q_seqlen))
+            out = flash_attn.flash_attn_varlen_qkvpacked_func(
+                qkv, cu_seqlens_q, max(q_seqlen)
+            )
         elif num_all_args == 2:
-            out = flash_attn.flash_attn_varlen_kvpacked_func(q, kv, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen))
+            out = flash_attn.flash_attn_varlen_kvpacked_func(
+                q, kv, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen)
+            )
         elif num_all_args == 3:
-            out = flash_attn.flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen))
+            out = flash_attn.flash_attn_varlen_func(
+                q, k, v, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen)
+            )
     else:
         raise ValueError(f"Unknown attention module: {ATTN}")
-    
+
     if s is not None:
         return s.replace(out)
     else:

trellis/modules/sparse/attention/modules.py

@@ -4,7 +4,10 @@ import torch.nn as nn
 import torch.nn.functional as F
 from .. import SparseTensor
 from .full_attn import sparse_scaled_dot_product_attention
-from .serialized_attn import SerializeMode, sparse_serialized_scaled_dot_product_self_attention
+from .serialized_attn import (
+    SerializeMode,
+    sparse_serialized_scaled_dot_product_self_attention,
+)
 from .windowed_attn import sparse_windowed_scaled_dot_product_self_attention
 from ...attention import RotaryPositionEmbedder
 
@@ -12,16 +15,18 @@ from ...attention import RotaryPositionEmbedder
 class SparseMultiHeadRMSNorm(nn.Module):
     def __init__(self, dim: int, heads: int):
         super().__init__()
-        self.scale = dim ** 0.5
+        self.scale = dim**0.5
         self.gamma = nn.Parameter(torch.ones(heads, dim))
 
-    def forward(self, x: Union[SparseTensor, torch.Tensor]) -> Union[SparseTensor, torch.Tensor]:
+    def forward(
+        self, x: Union[SparseTensor, torch.Tensor]
+    ) -> Union[SparseTensor, torch.Tensor]:
         x_type = x.dtype
         x = x.float()
         if isinstance(x, SparseTensor):
             x = x.replace(F.normalize(x.feats, dim=-1))
         else:
-            x = F.normalize(x, dim=-1)            
+            x = F.normalize(x, dim=-1)
         return (x * self.gamma * self.scale).to(x_type)
 
 
@@ -44,9 +49,17 @@ class SparseMultiHeadAttention(nn.Module):
         super().__init__()
         assert channels % num_heads == 0
         assert type in ["self", "cross"], f"Invalid attention type: {type}"
-        assert attn_mode in ["full", "serialized", "windowed"], f"Invalid attention mode: {attn_mode}"
-        assert type == "self" or attn_mode == "full", "Cross-attention only supports full attention"
-        assert type == "self" or use_rope is False, "Rotary position embeddings only supported for self-attention"
+        assert attn_mode in [
+            "full",
+            "serialized",
+            "windowed",
+        ], f"Invalid attention mode: {attn_mode}"
+        assert (
+            type == "self" or attn_mode == "full"
+        ), "Cross-attention only supports full attention"
+        assert (
+            type == "self" or use_rope is False
+        ), "Rotary position embeddings only supported for self-attention"
         self.channels = channels
         self.ctx_channels = ctx_channels if ctx_channels is not None else channels
         self.num_heads = num_heads
@@ -64,31 +77,37 @@ class SparseMultiHeadAttention(nn.Module):
         else:
             self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
             self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
-        
+
         if self.qk_rms_norm:
             self.q_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
             self.k_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
-            
+
         self.to_out = nn.Linear(channels, channels)
 
         if use_rope:
             self.rope = RotaryPositionEmbedder(channels)
 
     @staticmethod
-    def _linear(module: nn.Linear, x: Union[SparseTensor, torch.Tensor]) -> Union[SparseTensor, torch.Tensor]:
+    def _linear(
+        module: nn.Linear, x: Union[SparseTensor, torch.Tensor]
+    ) -> Union[SparseTensor, torch.Tensor]:
         if isinstance(x, SparseTensor):
             return x.replace(module(x.feats))
         else:
             return module(x)
 
     @staticmethod
-    def _reshape_chs(x: Union[SparseTensor, torch.Tensor], shape: Tuple[int, ...]) -> Union[SparseTensor, torch.Tensor]:
+    def _reshape_chs(
+        x: Union[SparseTensor, torch.Tensor], shape: Tuple[int, ...]
+    ) -> Union[SparseTensor, torch.Tensor]:
         if isinstance(x, SparseTensor):
             return x.reshape(*shape)
         else:
             return x.reshape(*x.shape[:2], *shape)
 
-    def _fused_pre(self, x: Union[SparseTensor, torch.Tensor], num_fused: int) -> Union[SparseTensor, torch.Tensor]:
+    def _fused_pre(
+        self, x: Union[SparseTensor, torch.Tensor], num_fused: int
+    ) -> Union[SparseTensor, torch.Tensor]:
         if isinstance(x, SparseTensor):
             x_feats = x.feats.unsqueeze(0)
         else:
@@ -97,12 +116,16 @@ class SparseMultiHeadAttention(nn.Module):
         return x.replace(x_feats.squeeze(0)) if isinstance(x, SparseTensor) else x_feats
 
     def _rope(self, qkv: SparseTensor) -> SparseTensor:
-        q, k, v = qkv.feats.unbind(dim=1)   # [T, H, C]
+        q, k, v = qkv.feats.unbind(dim=1)  # [T, H, C]
         q, k = self.rope(q, k, qkv.coords[:, 1:])
-        qkv = qkv.replace(torch.stack([q, k, v], dim=1)) 
+        qkv = qkv.replace(torch.stack([q, k, v], dim=1))
         return qkv
-    
-    def forward(self, x: Union[SparseTensor, torch.Tensor], context: Optional[Union[SparseTensor, torch.Tensor]] = None) -> Union[SparseTensor, torch.Tensor]:
+
+    def forward(
+        self,
+        x: Union[SparseTensor, torch.Tensor],
+        context: Optional[Union[SparseTensor, torch.Tensor]] = None,
+    ) -> Union[SparseTensor, torch.Tensor]:
         if self._type == "self":
             qkv = self._linear(self.to_qkv, x)
             qkv = self._fused_pre(qkv, num_fused=3)
@@ -117,7 +140,11 @@ class SparseMultiHeadAttention(nn.Module):
                 h = sparse_scaled_dot_product_attention(qkv)
             elif self.attn_mode == "serialized":
                 h = sparse_serialized_scaled_dot_product_self_attention(
-                    qkv, self.window_size, serialize_mode=self.serialize_mode, shift_sequence=self.shift_sequence, shift_window=self.shift_window
+                    qkv,
+                    self.window_size,
+                    serialize_mode=self.serialize_mode,
+                    shift_sequence=self.shift_sequence,
+                    shift_window=self.shift_window,
                 )
             elif self.attn_mode == "windowed":
                 h = sparse_windowed_scaled_dot_product_self_attention(

trellis/modules/sparse/attention/serialized_attn.py

@@ -5,16 +5,16 @@ import math
 from .. import SparseTensor
 from .. import DEBUG, ATTN
 
-if ATTN == 'xformers':
+if ATTN == "xformers":
     import xformers.ops as xops
-elif ATTN == 'flash_attn':
+elif ATTN == "flash_attn":
     import flash_attn
 else:
     raise ValueError(f"Unknown attention module: {ATTN}")
 
 
 __all__ = [
-    'sparse_serialized_scaled_dot_product_self_attention',
+    "sparse_serialized_scaled_dot_product_self_attention",
 ]
 
 
@@ -29,7 +29,7 @@ SerializeModes = [
     SerializeMode.Z_ORDER,
     SerializeMode.Z_ORDER_TRANSPOSED,
     SerializeMode.HILBERT,
-    SerializeMode.HILBERT_TRANSPOSED
+    SerializeMode.HILBERT_TRANSPOSED,
 ]
 
 
@@ -38,7 +38,7 @@ def calc_serialization(
     window_size: int,
     serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
     shift_sequence: int = 0,
-    shift_window: Tuple[int, int, int] = (0, 0, 0)
+    shift_window: Tuple[int, int, int] = (0, 0, 0),
 ) -> Tuple[torch.Tensor, torch.Tensor, List[int]]:
     """
     Calculate serialization and partitioning for a set of coordinates.
@@ -58,32 +58,38 @@ def calc_serialization(
     seq_lens = []
     seq_batch_indices = []
     offsets = [0]
-    
-    if 'vox2seq' not in globals():
+
+    if "vox2seq" not in globals():
         import vox2seq
 
     # Serialize the input
     serialize_coords = tensor.coords[:, 1:].clone()
-    serialize_coords += torch.tensor(shift_window, dtype=torch.int32, device=tensor.device).reshape(1, 3)
+    serialize_coords += torch.tensor(
+        shift_window, dtype=torch.int32, device=tensor.device
+    ).reshape(1, 3)
     if serialize_mode == SerializeMode.Z_ORDER:
-        code = vox2seq.encode(serialize_coords, mode='z_order', permute=[0, 1, 2])
+        code = vox2seq.encode(serialize_coords, mode="z_order", permute=[0, 1, 2])
     elif serialize_mode == SerializeMode.Z_ORDER_TRANSPOSED:
-        code = vox2seq.encode(serialize_coords, mode='z_order', permute=[1, 0, 2])
+        code = vox2seq.encode(serialize_coords, mode="z_order", permute=[1, 0, 2])
     elif serialize_mode == SerializeMode.HILBERT:
-        code = vox2seq.encode(serialize_coords, mode='hilbert', permute=[0, 1, 2])
+        code = vox2seq.encode(serialize_coords, mode="hilbert", permute=[0, 1, 2])
     elif serialize_mode == SerializeMode.HILBERT_TRANSPOSED:
-        code = vox2seq.encode(serialize_coords, mode='hilbert', permute=[1, 0, 2])
+        code = vox2seq.encode(serialize_coords, mode="hilbert", permute=[1, 0, 2])
     else:
         raise ValueError(f"Unknown serialize mode: {serialize_mode}")
-    
+
     for bi, s in enumerate(tensor.layout):
         num_points = s.stop - s.start
         num_windows = (num_points + window_size - 1) // window_size
         valid_window_size = num_points / num_windows
-        to_ordered = torch.argsort(code[s.start:s.stop])
+        to_ordered = torch.argsort(code[s.start : s.stop])
         if num_windows == 1:
             fwd_indices.append(to_ordered)
-            bwd_indices.append(torch.zeros_like(to_ordered).scatter_(0, to_ordered, torch.arange(num_points, device=tensor.device)))
+            bwd_indices.append(
+                torch.zeros_like(to_ordered).scatter_(
+                    0, to_ordered, torch.arange(num_points, device=tensor.device)
+                )
+            )
             fwd_indices[-1] += s.start
             bwd_indices[-1] += offsets[-1]
             seq_lens.append(num_points)
@@ -92,18 +98,39 @@ def calc_serialization(
         else:
             # Partition the input
             offset = 0
-            mids = [(i + 0.5) * valid_window_size + shift_sequence for i in range(num_windows)]
-            split = [math.floor(i * valid_window_size + shift_sequence) for i in range(num_windows + 1)]
-            bwd_index = torch.zeros((num_points,), dtype=torch.int64, device=tensor.device)
+            mids = [
+                (i + 0.5) * valid_window_size + shift_sequence
+                for i in range(num_windows)
+            ]
+            split = [
+                math.floor(i * valid_window_size + shift_sequence)
+                for i in range(num_windows + 1)
+            ]
+            bwd_index = torch.zeros(
+                (num_points,), dtype=torch.int64, device=tensor.device
+            )
             for i in range(num_windows):
                 mid = mids[i]
                 valid_start = split[i]
                 valid_end = split[i + 1]
                 padded_start = math.floor(mid - 0.5 * window_size)
                 padded_end = padded_start + window_size
-                fwd_indices.append(to_ordered[torch.arange(padded_start, padded_end, device=tensor.device) % num_points])
+                fwd_indices.append(
+                    to_ordered[
+                        torch.arange(padded_start, padded_end, device=tensor.device)
+                        % num_points
+                    ]
+                )
                 offset += valid_start - padded_start
-                bwd_index.scatter_(0, fwd_indices[-1][valid_start-padded_start:valid_end-padded_start], torch.arange(offset, offset + valid_end - valid_start, device=tensor.device))
+                bwd_index.scatter_(
+                    0,
+                    fwd_indices[-1][
+                        valid_start - padded_start : valid_end - padded_start
+                    ],
+                    torch.arange(
+                        offset, offset + valid_end - valid_start, device=tensor.device
+                    ),
+                )
                 offset += padded_end - valid_start
                 fwd_indices[-1] += s.start
             seq_lens.extend([window_size] * num_windows)
@@ -115,14 +142,14 @@ def calc_serialization(
     bwd_indices = torch.cat(bwd_indices)
 
     return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
-    
+
 
 def sparse_serialized_scaled_dot_product_self_attention(
     qkv: SparseTensor,
     window_size: int,
     serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
     shift_sequence: int = 0,
-    shift_window: Tuple[int, int, int] = (0, 0, 0)
+    shift_window: Tuple[int, int, int] = (0, 0, 0),
 ) -> SparseTensor:
     """
     Apply serialized scaled dot product self attention to a sparse tensor.
@@ -135,59 +162,86 @@ def sparse_serialized_scaled_dot_product_self_attention(
         shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
         shift (int): The shift to use.
     """
-    assert len(qkv.shape) == 4 and qkv.shape[1] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
-
-    serialization_spatial_cache_name = f'serialization_{serialize_mode}_{window_size}_{shift_sequence}_{shift_window}'
-    serialization_spatial_cache = qkv.get_spatial_cache(serialization_spatial_cache_name)
+    assert (
+        len(qkv.shape) == 4 and qkv.shape[1] == 3
+    ), f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
+
+    serialization_spatial_cache_name = (
+        f"serialization_{serialize_mode}_{window_size}_{shift_sequence}_{shift_window}"
+    )
+    serialization_spatial_cache = qkv.get_spatial_cache(
+        serialization_spatial_cache_name
+    )
     if serialization_spatial_cache is None:
-        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_serialization(qkv, window_size, serialize_mode, shift_sequence, shift_window)
-        qkv.register_spatial_cache(serialization_spatial_cache_name, (fwd_indices, bwd_indices, seq_lens, seq_batch_indices))
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_serialization(
+            qkv, window_size, serialize_mode, shift_sequence, shift_window
+        )
+        qkv.register_spatial_cache(
+            serialization_spatial_cache_name,
+            (fwd_indices, bwd_indices, seq_lens, seq_batch_indices),
+        )
     else:
-        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = serialization_spatial_cache
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = (
+            serialization_spatial_cache
+        )
 
     M = fwd_indices.shape[0]
     T = qkv.feats.shape[0]
     H = qkv.feats.shape[2]
     C = qkv.feats.shape[3]
-    
-    qkv_feats = qkv.feats[fwd_indices]      # [M, 3, H, C]
+
+    qkv_feats = qkv.feats[fwd_indices]  # [M, 3, H, C]
 
     if DEBUG:
         start = 0
         qkv_coords = qkv.coords[fwd_indices]
         for i in range(len(seq_lens)):
-            assert (qkv_coords[start:start+seq_lens[i], 0] == seq_batch_indices[i]).all(), f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
+            assert (
+                qkv_coords[start : start + seq_lens[i], 0] == seq_batch_indices[i]
+            ).all(), (
+                f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
+            )
             start += seq_lens[i]
 
     if all([seq_len == window_size for seq_len in seq_lens]):
         B = len(seq_lens)
         N = window_size
         qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
-        if ATTN == 'xformers':
-            q, k, v = qkv_feats.unbind(dim=2)                       # [B, N, H, C]
-            out = xops.memory_efficient_attention(q, k, v)          # [B, N, H, C]
-        elif ATTN == 'flash_attn':
-            out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)   # [B, N, H, C]
+        if ATTN == "xformers":
+            q, k, v = qkv_feats.unbind(dim=2)  # [B, N, H, C]
+            out = xops.memory_efficient_attention(q, k, v)  # [B, N, H, C]
+        elif ATTN == "flash_attn":
+            out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)  # [B, N, H, C]
         else:
             raise ValueError(f"Unknown attention module: {ATTN}")
-        out = out.reshape(B * N, H, C)                              # [M, H, C]
+        out = out.reshape(B * N, H, C)  # [M, H, C]
     else:
-        if ATTN == 'xformers':
-            q, k, v = qkv_feats.unbind(dim=1)                       # [M, H, C]
-            q = q.unsqueeze(0)                                      # [1, M, H, C]
-            k = k.unsqueeze(0)                                      # [1, M, H, C]
-            v = v.unsqueeze(0)                                      # [1, M, H, C]
+        if ATTN == "xformers":
+            q, k, v = qkv_feats.unbind(dim=1)  # [M, H, C]
+            q = q.unsqueeze(0)  # [1, M, H, C]
+            k = k.unsqueeze(0)  # [1, M, H, C]
+            v = v.unsqueeze(0)  # [1, M, H, C]
             mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
-            out = xops.memory_efficient_attention(q, k, v, mask)[0] # [M, H, C]
-        elif ATTN == 'flash_attn':
-            cu_seqlens = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)], dim=0) \
-                        .to(qkv.device).int()
-            out = flash_attn.flash_attn_varlen_qkvpacked_func(qkv_feats, cu_seqlens, max(seq_lens)) # [M, H, C]
-
-    out = out[bwd_indices]      # [T, H, C]
+            out = xops.memory_efficient_attention(q, k, v, mask)[0]  # [M, H, C]
+        elif ATTN == "flash_attn":
+            cu_seqlens = (
+                torch.cat(
+                    [torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)],
+                    dim=0,
+                )
+                .to(qkv.device)
+                .int()
+            )
+            out = flash_attn.flash_attn_varlen_qkvpacked_func(
+                qkv_feats, cu_seqlens, max(seq_lens)
+            )  # [M, H, C]
+
+    out = out[bwd_indices]  # [T, H, C]
 
     if DEBUG:
         qkv_coords = qkv_coords[bwd_indices]
-        assert torch.equal(qkv_coords, qkv.coords), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
+        assert torch.equal(
+            qkv_coords, qkv.coords
+        ), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
 
     return qkv.replace(out)

trellis/modules/sparse/attention/windowed_attn.py

@@ -4,23 +4,23 @@ import math
 from .. import SparseTensor
 from .. import DEBUG, ATTN
 
-if ATTN == 'xformers':
+if ATTN == "xformers":
     import xformers.ops as xops
-elif ATTN == 'flash_attn':
+elif ATTN == "flash_attn":
     import flash_attn
 else:
     raise ValueError(f"Unknown attention module: {ATTN}")
 
 
 __all__ = [
-    'sparse_windowed_scaled_dot_product_self_attention',
+    "sparse_windowed_scaled_dot_product_self_attention",
 ]
 
 
 def calc_window_partition(
     tensor: SparseTensor,
     window_size: Union[int, Tuple[int, ...]],
-    shift_window: Union[int, Tuple[int, ...]] = 0
+    shift_window: Union[int, Tuple[int, ...]] = 0,
 ) -> Tuple[torch.Tensor, torch.Tensor, List[int], List[int]]:
     """
     Calculate serialization and partitioning for a set of coordinates.
@@ -37,33 +37,43 @@ def calc_window_partition(
         (List[int]): Sequence batch indices.
     """
     DIM = tensor.coords.shape[1] - 1
-    shift_window = (shift_window,) * DIM if isinstance(shift_window, int) else shift_window
+    shift_window = (
+        (shift_window,) * DIM if isinstance(shift_window, int) else shift_window
+    )
     window_size = (window_size,) * DIM if isinstance(window_size, int) else window_size
     shifted_coords = tensor.coords.clone().detach()
-    shifted_coords[:, 1:] += torch.tensor(shift_window, device=tensor.device, dtype=torch.int32).unsqueeze(0)
+    shifted_coords[:, 1:] += torch.tensor(
+        shift_window, device=tensor.device, dtype=torch.int32
+    ).unsqueeze(0)
 
     MAX_COORDS = shifted_coords[:, 1:].max(dim=0).values.tolist()
     NUM_WINDOWS = [math.ceil((mc + 1) / ws) for mc, ws in zip(MAX_COORDS, window_size)]
     OFFSET = torch.cumprod(torch.tensor([1] + NUM_WINDOWS[::-1]), dim=0).tolist()[::-1]
 
-    shifted_coords[:, 1:] //= torch.tensor(window_size, device=tensor.device, dtype=torch.int32).unsqueeze(0)
-    shifted_indices = (shifted_coords * torch.tensor(OFFSET, device=tensor.device, dtype=torch.int32).unsqueeze(0)).sum(dim=1)
+    shifted_coords[:, 1:] //= torch.tensor(
+        window_size, device=tensor.device, dtype=torch.int32
+    ).unsqueeze(0)
+    shifted_indices = (
+        shifted_coords
+        * torch.tensor(OFFSET, device=tensor.device, dtype=torch.int32).unsqueeze(0)
+    ).sum(dim=1)
     fwd_indices = torch.argsort(shifted_indices)
     bwd_indices = torch.empty_like(fwd_indices)
     bwd_indices[fwd_indices] = torch.arange(fwd_indices.shape[0], device=tensor.device)
     seq_lens = torch.bincount(shifted_indices)
-    seq_batch_indices = torch.arange(seq_lens.shape[0], device=tensor.device, dtype=torch.int32) // OFFSET[0]
+    seq_batch_indices = (
+        torch.arange(seq_lens.shape[0], device=tensor.device, dtype=torch.int32)
+        // OFFSET[0]
+    )
     mask = seq_lens != 0
     seq_lens = seq_lens[mask].tolist()
     seq_batch_indices = seq_batch_indices[mask].tolist()
 
     return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
-    
+
 
 def sparse_windowed_scaled_dot_product_self_attention(
-    qkv: SparseTensor,
-    window_size: int,
-    shift_window: Tuple[int, int, int] = (0, 0, 0)
+    qkv: SparseTensor, window_size: int, shift_window: Tuple[int, int, int] = (0, 0, 0)
 ) -> SparseTensor:
     """
     Apply windowed scaled dot product self attention to a sparse tensor.
@@ -74,62 +84,92 @@ def sparse_windowed_scaled_dot_product_self_attention(
         shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
         shift (int): The shift to use.
     """
-    assert len(qkv.shape) == 4 and qkv.shape[1] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
-
-    serialization_spatial_cache_name = f'window_partition_{window_size}_{shift_window}'
-    serialization_spatial_cache = qkv.get_spatial_cache(serialization_spatial_cache_name)
+    assert (
+        len(qkv.shape) == 4 and qkv.shape[1] == 3
+    ), f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
+
+    serialization_spatial_cache_name = f"window_partition_{window_size}_{shift_window}"
+    serialization_spatial_cache = qkv.get_spatial_cache(
+        serialization_spatial_cache_name
+    )
     if serialization_spatial_cache is None:
-        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_window_partition(qkv, window_size, shift_window)
-        qkv.register_spatial_cache(serialization_spatial_cache_name, (fwd_indices, bwd_indices, seq_lens, seq_batch_indices))
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_window_partition(
+            qkv, window_size, shift_window
+        )
+        qkv.register_spatial_cache(
+            serialization_spatial_cache_name,
+            (fwd_indices, bwd_indices, seq_lens, seq_batch_indices),
+        )
     else:
-        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = serialization_spatial_cache
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = (
+            serialization_spatial_cache
+        )
 
     M = fwd_indices.shape[0]
     T = qkv.feats.shape[0]
     H = qkv.feats.shape[2]
     C = qkv.feats.shape[3]
-    
-    qkv_feats = qkv.feats[fwd_indices]      # [M, 3, H, C]
+
+    qkv_feats = qkv.feats[fwd_indices]  # [M, 3, H, C]
 
     if DEBUG:
         start = 0
         qkv_coords = qkv.coords[fwd_indices]
         for i in range(len(seq_lens)):
-            seq_coords = qkv_coords[start:start+seq_lens[i]]
-            assert (seq_coords[:, 0] == seq_batch_indices[i]).all(), f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
-            assert (seq_coords[:, 1:].max(dim=0).values - seq_coords[:, 1:].min(dim=0).values < window_size).all(), \
-                    f"SparseWindowedScaledDotProductSelfAttention: window size exceeded"
+            seq_coords = qkv_coords[start : start + seq_lens[i]]
+            assert (
+                seq_coords[:, 0] == seq_batch_indices[i]
+            ).all(), (
+                f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
+            )
+            assert (
+                seq_coords[:, 1:].max(dim=0).values
+                - seq_coords[:, 1:].min(dim=0).values
+                < window_size
+            ).all(), (
+                f"SparseWindowedScaledDotProductSelfAttention: window size exceeded"
+            )
             start += seq_lens[i]
 
     if all([seq_len == window_size for seq_len in seq_lens]):
         B = len(seq_lens)
         N = window_size
         qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
-        if ATTN == 'xformers':
-            q, k, v = qkv_feats.unbind(dim=2)                       # [B, N, H, C]
-            out = xops.memory_efficient_attention(q, k, v)          # [B, N, H, C]
-        elif ATTN == 'flash_attn':
-            out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)   # [B, N, H, C]
+        if ATTN == "xformers":
+            q, k, v = qkv_feats.unbind(dim=2)  # [B, N, H, C]
+            out = xops.memory_efficient_attention(q, k, v)  # [B, N, H, C]
+        elif ATTN == "flash_attn":
+            out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)  # [B, N, H, C]
         else:
             raise ValueError(f"Unknown attention module: {ATTN}")
-        out = out.reshape(B * N, H, C)                              # [M, H, C]
+        out = out.reshape(B * N, H, C)  # [M, H, C]
     else:
-        if ATTN == 'xformers':
-            q, k, v = qkv_feats.unbind(dim=1)                       # [M, H, C]
-            q = q.unsqueeze(0)                                      # [1, M, H, C]
-            k = k.unsqueeze(0)                                      # [1, M, H, C]
-            v = v.unsqueeze(0)                                      # [1, M, H, C]
+        if ATTN == "xformers":
+            q, k, v = qkv_feats.unbind(dim=1)  # [M, H, C]
+            q = q.unsqueeze(0)  # [1, M, H, C]
+            k = k.unsqueeze(0)  # [1, M, H, C]
+            v = v.unsqueeze(0)  # [1, M, H, C]
             mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
-            out = xops.memory_efficient_attention(q, k, v, mask)[0] # [M, H, C]
-        elif ATTN == 'flash_attn':
-            cu_seqlens = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)], dim=0) \
-                        .to(qkv.device).int()
-            out = flash_attn.flash_attn_varlen_qkvpacked_func(qkv_feats, cu_seqlens, max(seq_lens)) # [M, H, C]
-
-    out = out[bwd_indices]      # [T, H, C]
+            out = xops.memory_efficient_attention(q, k, v, mask)[0]  # [M, H, C]
+        elif ATTN == "flash_attn":
+            cu_seqlens = (
+                torch.cat(
+                    [torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)],
+                    dim=0,
+                )
+                .to(qkv.device)
+                .int()
+            )
+            out = flash_attn.flash_attn_varlen_qkvpacked_func(
+                qkv_feats, cu_seqlens, max(seq_lens)
+            )  # [M, H, C]
+
+    out = out[bwd_indices]  # [T, H, C]
 
     if DEBUG:
         qkv_coords = qkv_coords[bwd_indices]
-        assert torch.equal(qkv_coords, qkv.coords), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
+        assert torch.equal(
+            qkv_coords, qkv.coords
+        ), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
 
     return qkv.replace(out)

trellis/modules/sparse/basic.py

@@ -2,22 +2,23 @@ from typing import *
 import torch
 import torch.nn as nn
 from . import BACKEND, DEBUG
-SparseTensorData = None # Lazy import
+
+SparseTensorData = None  # Lazy import
 
 
 __all__ = [
-    'SparseTensor',
-    'sparse_batch_broadcast',
-    'sparse_batch_op',
-    'sparse_cat',
-    'sparse_unbind',
+    "SparseTensor",
+    "sparse_batch_broadcast",
+    "sparse_batch_op",
+    "sparse_cat",
+    "sparse_unbind",
 ]
 
 
 class SparseTensor:
     """
     Sparse tensor with support for both torchsparse and spconv backends.
-    
+
     Parameters:
     - feats (torch.Tensor): Features of the sparse tensor.
     - coords (torch.Tensor): Coordinates of the sparse tensor.
@@ -29,64 +30,87 @@ class SparseTensor:
     - Data corresponding to a same batch should be contiguous.
     - Coords should be in [0, 1023]
     """
+
     @overload
-    def __init__(self, feats: torch.Tensor, coords: torch.Tensor, shape: Optional[torch.Size] = None, layout: Optional[List[slice]] = None, **kwargs): ...
+    def __init__(
+        self,
+        feats: torch.Tensor,
+        coords: torch.Tensor,
+        shape: Optional[torch.Size] = None,
+        layout: Optional[List[slice]] = None,
+        **kwargs,
+    ): ...
 
     @overload
-    def __init__(self, data, shape: Optional[torch.Size] = None, layout: Optional[List[slice]] = None, **kwargs): ...
+    def __init__(
+        self,
+        data,
+        shape: Optional[torch.Size] = None,
+        layout: Optional[List[slice]] = None,
+        **kwargs,
+    ): ...
 
     def __init__(self, *args, **kwargs):
         # Lazy import of sparse tensor backend
         global SparseTensorData
         if SparseTensorData is None:
             import importlib
-            if BACKEND == 'torchsparse':
-                SparseTensorData = importlib.import_module('torchsparse').SparseTensor
-            elif BACKEND == 'spconv':
-                SparseTensorData = importlib.import_module('spconv.pytorch').SparseConvTensor
-                
+
+            if BACKEND == "torchsparse":
+                SparseTensorData = importlib.import_module("torchsparse").SparseTensor
+            elif BACKEND == "spconv":
+                SparseTensorData = importlib.import_module(
+                    "spconv.pytorch"
+                ).SparseConvTensor
+
         method_id = 0
         if len(args) != 0:
             method_id = 0 if isinstance(args[0], torch.Tensor) else 1
         else:
-            method_id = 1 if 'data' in kwargs else 0
+            method_id = 1 if "data" in kwargs else 0
 
         if method_id == 0:
             feats, coords, shape, layout = args + (None,) * (4 - len(args))
-            if 'feats' in kwargs:
-                feats = kwargs['feats']
-                del kwargs['feats']
-            if 'coords' in kwargs:
-                coords = kwargs['coords']
-                del kwargs['coords']
-            if 'shape' in kwargs:
-                shape = kwargs['shape']
-                del kwargs['shape']
-            if 'layout' in kwargs:
-                layout = kwargs['layout']
-                del kwargs['layout']
+            if "feats" in kwargs:
+                feats = kwargs["feats"]
+                del kwargs["feats"]
+            if "coords" in kwargs:
+                coords = kwargs["coords"]
+                del kwargs["coords"]
+            if "shape" in kwargs:
+                shape = kwargs["shape"]
+                del kwargs["shape"]
+            if "layout" in kwargs:
+                layout = kwargs["layout"]
+                del kwargs["layout"]
 
             if shape is None:
                 shape = self.__cal_shape(feats, coords)
             if layout is None:
                 layout = self.__cal_layout(coords, shape[0])
-            if BACKEND == 'torchsparse':
+            if BACKEND == "torchsparse":
                 self.data = SparseTensorData(feats, coords, **kwargs)
-            elif BACKEND == 'spconv':
+            elif BACKEND == "spconv":
                 spatial_shape = list(coords.max(0)[0] + 1)[1:]
-                self.data = SparseTensorData(feats.reshape(feats.shape[0], -1), coords, spatial_shape, shape[0], **kwargs)
+                self.data = SparseTensorData(
+                    feats.reshape(feats.shape[0], -1),
+                    coords,
+                    spatial_shape,
+                    shape[0],
+                    **kwargs,
+                )
                 self.data._features = feats
         elif method_id == 1:
             data, shape, layout = args + (None,) * (3 - len(args))
-            if 'data' in kwargs:
-                data = kwargs['data']
-                del kwargs['data']
-            if 'shape' in kwargs:
-                shape = kwargs['shape']
-                del kwargs['shape']
-            if 'layout' in kwargs:
-                layout = kwargs['layout']
-                del kwargs['layout']
+            if "data" in kwargs:
+                data = kwargs["data"]
+                del kwargs["data"]
+            if "shape" in kwargs:
+                shape = kwargs["shape"]
+                del kwargs["shape"]
+            if "layout" in kwargs:
+                layout = kwargs["layout"]
+                del kwargs["layout"]
 
             self.data = data
             if shape is None:
@@ -96,73 +120,84 @@ class SparseTensor:
 
         self._shape = shape
         self._layout = layout
-        self._scale = kwargs.get('scale', (1, 1, 1))
-        self._spatial_cache = kwargs.get('spatial_cache', {})
+        self._scale = kwargs.get("scale", (1, 1, 1))
+        self._spatial_cache = kwargs.get("spatial_cache", {})
 
         if DEBUG:
             try:
-                assert self.feats.shape[0] == self.coords.shape[0], f"Invalid feats shape: {self.feats.shape}, coords shape: {self.coords.shape}"
-                assert self.shape == self.__cal_shape(self.feats, self.coords), f"Invalid shape: {self.shape}"
-                assert self.layout == self.__cal_layout(self.coords, self.shape[0]), f"Invalid layout: {self.layout}"
+                assert (
+                    self.feats.shape[0] == self.coords.shape[0]
+                ), f"Invalid feats shape: {self.feats.shape}, coords shape: {self.coords.shape}"
+                assert self.shape == self.__cal_shape(
+                    self.feats, self.coords
+                ), f"Invalid shape: {self.shape}"
+                assert self.layout == self.__cal_layout(
+                    self.coords, self.shape[0]
+                ), f"Invalid layout: {self.layout}"
                 for i in range(self.shape[0]):
-                    assert torch.all(self.coords[self.layout[i], 0] == i), f"The data of batch {i} is not contiguous"
+                    assert torch.all(
+                        self.coords[self.layout[i], 0] == i
+                    ), f"The data of batch {i} is not contiguous"
             except Exception as e:
-                print('Debugging information:')
+                print("Debugging information:")
                 print(f"- Shape: {self.shape}")
                 print(f"- Layout: {self.layout}")
                 print(f"- Scale: {self._scale}")
                 print(f"- Coords: {self.coords}")
                 raise e
-        
+
     def __cal_shape(self, feats, coords):
         shape = []
         shape.append(coords[:, 0].max().item() + 1)
         shape.extend([*feats.shape[1:]])
         return torch.Size(shape)
-    
+
     def __cal_layout(self, coords, batch_size):
         seq_len = torch.bincount(coords[:, 0], minlength=batch_size)
-        offset = torch.cumsum(seq_len, dim=0) 
-        layout = [slice((offset[i] - seq_len[i]).item(), offset[i].item()) for i in range(batch_size)]
+        offset = torch.cumsum(seq_len, dim=0)
+        layout = [
+            slice((offset[i] - seq_len[i]).item(), offset[i].item())
+            for i in range(batch_size)
+        ]
         return layout
-    
+
     @property
     def shape(self) -> torch.Size:
         return self._shape
-    
+
     def dim(self) -> int:
         return len(self.shape)
-    
+
     @property
     def layout(self) -> List[slice]:
         return self._layout
 
     @property
     def feats(self) -> torch.Tensor:
-        if BACKEND == 'torchsparse':
+        if BACKEND == "torchsparse":
             return self.data.F
-        elif BACKEND == 'spconv':
+        elif BACKEND == "spconv":
             return self.data.features
-    
+
     @feats.setter
     def feats(self, value: torch.Tensor):
-        if BACKEND == 'torchsparse':
+        if BACKEND == "torchsparse":
             self.data.F = value
-        elif BACKEND == 'spconv':
+        elif BACKEND == "spconv":
             self.data.features = value
 
     @property
     def coords(self) -> torch.Tensor:
-        if BACKEND == 'torchsparse':
+        if BACKEND == "torchsparse":
             return self.data.C
-        elif BACKEND == 'spconv':
+        elif BACKEND == "spconv":
             return self.data.indices
-        
+
     @coords.setter
     def coords(self, value: torch.Tensor):
-        if BACKEND == 'torchsparse':
+        if BACKEND == "torchsparse":
             self.data.C = value
-        elif BACKEND == 'spconv':
+        elif BACKEND == "spconv":
             self.data.indices = value
 
     @property
@@ -174,12 +209,16 @@ class SparseTensor:
         return self.feats.device
 
     @overload
-    def to(self, dtype: torch.dtype) -> 'SparseTensor': ...
+    def to(self, dtype: torch.dtype) -> "SparseTensor": ...
 
     @overload
-    def to(self, device: Optional[Union[str, torch.device]] = None, dtype: Optional[torch.dtype] = None) -> 'SparseTensor': ...
+    def to(
+        self,
+        device: Optional[Union[str, torch.device]] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> "SparseTensor": ...
 
-    def to(self, *args, **kwargs) -> 'SparseTensor':
+    def to(self, *args, **kwargs) -> "SparseTensor":
         device = None
         dtype = None
         if len(args) == 2:
@@ -189,13 +228,13 @@ class SparseTensor:
                 dtype = args[0]
             else:
                 device = args[0]
-        if 'dtype' in kwargs:
+        if "dtype" in kwargs:
             assert dtype is None, "to() received multiple values for argument 'dtype'"
-            dtype = kwargs['dtype']
-        if 'device' in kwargs:
+            dtype = kwargs["dtype"]
+        if "device" in kwargs:
             assert device is None, "to() received multiple values for argument 'device'"
-            device = kwargs['device']
-        
+            device = kwargs["device"]
+
         new_feats = self.feats.to(device=device, dtype=dtype)
         new_coords = self.coords.to(device=device)
         return self.replace(new_feats, new_coords)
@@ -204,46 +243,48 @@ class SparseTensor:
         new_feats = self.feats.type(dtype)
         return self.replace(new_feats)
 
-    def cpu(self) -> 'SparseTensor':
+    def cpu(self) -> "SparseTensor":
         new_feats = self.feats.cpu()
         new_coords = self.coords.cpu()
         return self.replace(new_feats, new_coords)
-    
-    def cuda(self) -> 'SparseTensor':
+
+    def cuda(self) -> "SparseTensor":
         new_feats = self.feats.cuda()
         new_coords = self.coords.cuda()
         return self.replace(new_feats, new_coords)
 
-    def half(self) -> 'SparseTensor':
+    def half(self) -> "SparseTensor":
         new_feats = self.feats.half()
         return self.replace(new_feats)
-    
-    def float(self) -> 'SparseTensor':
+
+    def float(self) -> "SparseTensor":
         new_feats = self.feats.float()
         return self.replace(new_feats)
-    
-    def detach(self) -> 'SparseTensor':
+
+    def detach(self) -> "SparseTensor":
         new_coords = self.coords.detach()
         new_feats = self.feats.detach()
         return self.replace(new_feats, new_coords)
 
     def dense(self) -> torch.Tensor:
-        if BACKEND == 'torchsparse':
+        if BACKEND == "torchsparse":
             return self.data.dense()
-        elif BACKEND == 'spconv':
+        elif BACKEND == "spconv":
             return self.data.dense()
 
-    def reshape(self, *shape) -> 'SparseTensor':
+    def reshape(self, *shape) -> "SparseTensor":
         new_feats = self.feats.reshape(self.feats.shape[0], *shape)
         return self.replace(new_feats)
-    
-    def unbind(self, dim: int) -> List['SparseTensor']:
+
+    def unbind(self, dim: int) -> List["SparseTensor"]:
         return sparse_unbind(self, dim)
 
-    def replace(self, feats: torch.Tensor, coords: Optional[torch.Tensor] = None) -> 'SparseTensor':
+    def replace(
+        self, feats: torch.Tensor, coords: Optional[torch.Tensor] = None
+    ) -> "SparseTensor":
         new_shape = [self.shape[0]]
         new_shape.extend(feats.shape[1:])
-        if BACKEND == 'torchsparse':
+        if BACKEND == "torchsparse":
             new_data = SparseTensorData(
                 feats=feats,
                 coords=self.data.coords if coords is None else coords,
@@ -251,7 +292,7 @@ class SparseTensor:
                 spatial_range=self.data.spatial_range,
             )
             new_data._caches = self.data._caches
-        elif BACKEND == 'spconv':
+        elif BACKEND == "spconv":
             new_data = SparseTensorData(
                 self.data.features.reshape(self.data.features.shape[0], -1),
                 self.data.indices,
@@ -259,7 +300,7 @@ class SparseTensor:
                 self.data.batch_size,
                 self.data.grid,
                 self.data.voxel_num,
-                self.data.indice_dict
+                self.data.indice_dict,
             )
             new_data._features = feats
             new_data.benchmark = self.data.benchmark
@@ -270,26 +311,39 @@ class SparseTensor:
             new_data.int8_scale = self.data.int8_scale
             if coords is not None:
                 new_data.indices = coords
-        new_tensor = SparseTensor(new_data, shape=torch.Size(new_shape), layout=self.layout, scale=self._scale, spatial_cache=self._spatial_cache)
+        new_tensor = SparseTensor(
+            new_data,
+            shape=torch.Size(new_shape),
+            layout=self.layout,
+            scale=self._scale,
+            spatial_cache=self._spatial_cache,
+        )
         return new_tensor
 
     @staticmethod
-    def full(aabb, dim, value, dtype=torch.float32, device=None) -> 'SparseTensor':
+    def full(aabb, dim, value, dtype=torch.float32, device=None) -> "SparseTensor":
         N, C = dim
         x = torch.arange(aabb[0], aabb[3] + 1)
         y = torch.arange(aabb[1], aabb[4] + 1)
         z = torch.arange(aabb[2], aabb[5] + 1)
-        coords = torch.stack(torch.meshgrid(x, y, z, indexing='ij'), dim=-1).reshape(-1, 3)
-        coords = torch.cat([
-            torch.arange(N).view(-1, 1).repeat(1, coords.shape[0]).view(-1, 1),
-            coords.repeat(N, 1),
-        ], dim=1).to(dtype=torch.int32, device=device)
+        coords = torch.stack(torch.meshgrid(x, y, z, indexing="ij"), dim=-1).reshape(
+            -1, 3
+        )
+        coords = torch.cat(
+            [
+                torch.arange(N).view(-1, 1).repeat(1, coords.shape[0]).view(-1, 1),
+                coords.repeat(N, 1),
+            ],
+            dim=1,
+        ).to(dtype=torch.int32, device=device)
         feats = torch.full((coords.shape[0], C), value, dtype=dtype, device=device)
         return SparseTensor(feats=feats, coords=coords)
 
-    def __merge_sparse_cache(self, other: 'SparseTensor') -> dict:
+    def __merge_sparse_cache(self, other: "SparseTensor") -> dict:
         new_cache = {}
-        for k in set(list(self._spatial_cache.keys()) + list(other._spatial_cache.keys())):
+        for k in set(
+            list(self._spatial_cache.keys()) + list(other._spatial_cache.keys())
+        ):
             if k in self._spatial_cache:
                 new_cache[k] = self._spatial_cache[k]
             if k in other._spatial_cache:
@@ -299,10 +353,12 @@ class SparseTensor:
                     new_cache[k].update(other._spatial_cache[k])
         return new_cache
 
-    def __neg__(self) -> 'SparseTensor':
+    def __neg__(self) -> "SparseTensor":
         return self.replace(-self.feats)
-    
-    def __elemwise__(self, other: Union[torch.Tensor, 'SparseTensor'], op: callable) -> 'SparseTensor':
+
+    def __elemwise__(
+        self, other: Union[torch.Tensor, "SparseTensor"], op: callable
+    ) -> "SparseTensor":
         if isinstance(other, torch.Tensor):
             try:
                 other = torch.broadcast_to(other, self.shape)
@@ -317,28 +373,44 @@ class SparseTensor:
             new_tensor._spatial_cache = self.__merge_sparse_cache(other)
         return new_tensor
 
-    def __add__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+    def __add__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, torch.add)
 
-    def __radd__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+    def __radd__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, torch.add)
-    
-    def __sub__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+
+    def __sub__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, torch.sub)
-    
-    def __rsub__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+
+    def __rsub__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, lambda x, y: torch.sub(y, x))
 
-    def __mul__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+    def __mul__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, torch.mul)
 
-    def __rmul__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+    def __rmul__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, torch.mul)
 
-    def __truediv__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+    def __truediv__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, torch.div)
 
-    def __rtruediv__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+    def __rtruediv__(
+        self, other: Union[torch.Tensor, "SparseTensor", float]
+    ) -> "SparseTensor":
         return self.__elemwise__(other, lambda x, y: torch.div(y, x))
 
     def __getitem__(self, idx):
@@ -348,7 +420,9 @@ class SparseTensor:
             idx = range(*idx.indices(self.shape[0]))
         elif isinstance(idx, torch.Tensor):
             if idx.dtype == torch.bool:
-                assert idx.shape == (self.shape[0],), f"Invalid index shape: {idx.shape}"
+                assert idx.shape == (
+                    self.shape[0],
+                ), f"Invalid index shape: {idx.shape}"
                 idx = idx.nonzero().squeeze(1)
             elif idx.dtype in [torch.int32, torch.int64]:
                 assert len(idx.shape) == 1, f"Invalid index shape: {idx.shape}"
@@ -356,7 +430,7 @@ class SparseTensor:
                 raise ValueError(f"Unknown index type: {idx.dtype}")
         else:
             raise ValueError(f"Unknown index type: {type(idx)}")
-        
+
         coords = []
         feats = []
         for new_idx, old_idx in enumerate(idx):
@@ -392,7 +466,7 @@ class SparseTensor:
 def sparse_batch_broadcast(input: SparseTensor, other: torch.Tensor) -> torch.Tensor:
     """
     Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
-    
+
     Args:
         input (torch.Tensor): 1D tensor to broadcast.
         target (SparseTensor): Sparse tensor to broadcast to.
@@ -405,10 +479,12 @@ def sparse_batch_broadcast(input: SparseTensor, other: torch.Tensor) -> torch.Te
     return broadcasted
 
 
-def sparse_batch_op(input: SparseTensor, other: torch.Tensor, op: callable = torch.add) -> SparseTensor:
+def sparse_batch_op(
+    input: SparseTensor, other: torch.Tensor, op: callable = torch.add
+) -> SparseTensor:
     """
     Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
-    
+
     Args:
         input (torch.Tensor): 1D tensor to broadcast.
         target (SparseTensor): Sparse tensor to broadcast to.
@@ -420,7 +496,7 @@ def sparse_batch_op(input: SparseTensor, other: torch.Tensor, op: callable = tor
 def sparse_cat(inputs: List[SparseTensor], dim: int = 0) -> SparseTensor:
     """
     Concatenate a list of sparse tensors.
-    
+
     Args:
         inputs (List[SparseTensor]): List of sparse tensors to concatenate.
     """
@@ -447,7 +523,7 @@ def sparse_cat(inputs: List[SparseTensor], dim: int = 0) -> SparseTensor:
 def sparse_unbind(input: SparseTensor, dim: int) -> List[SparseTensor]:
     """
     Unbind a sparse tensor along a dimension.
-    
+
     Args:
         input (SparseTensor): Sparse tensor to unbind.
         dim (int): Dimension to unbind.

trellis/modules/sparse/conv/__init__.py

@@ -1,21 +1,26 @@
 from .. import BACKEND
 
 
-SPCONV_ALGO = 'auto'    # 'auto', 'implicit_gemm', 'native'
+SPCONV_ALGO = "auto"  # 'auto', 'implicit_gemm', 'native'
+
 
 def __from_env():
     import os
-        
+
     global SPCONV_ALGO
-    env_spconv_algo = os.environ.get('SPCONV_ALGO')
-    if env_spconv_algo is not None and env_spconv_algo in ['auto', 'implicit_gemm', 'native']:
+    env_spconv_algo = os.environ.get("SPCONV_ALGO")
+    if env_spconv_algo is not None and env_spconv_algo in [
+        "auto",
+        "implicit_gemm",
+        "native",
+    ]:
         SPCONV_ALGO = env_spconv_algo
     print(f"[SPARSE][CONV] spconv algo: {SPCONV_ALGO}")
-        
+
 
 __from_env()
 
-if BACKEND == 'torchsparse':
+if BACKEND == "torchsparse":
     from .conv_torchsparse import *
-elif BACKEND == 'spconv':
+elif BACKEND == "spconv":
     from .conv_spconv import *

trellis/modules/sparse/conv/conv_spconv.py

@@ -4,21 +4,54 @@ from .. import SparseTensor
 from .. import DEBUG
 from . import SPCONV_ALGO
 
+
 class SparseConv3d(nn.Module):
-    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, padding=None, bias=True, indice_key=None):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        dilation=1,
+        padding=None,
+        bias=True,
+        indice_key=None,
+    ):
         super(SparseConv3d, self).__init__()
-        if 'spconv' not in globals():
+        if "spconv" not in globals():
             import spconv.pytorch as spconv
         algo = None
-        if SPCONV_ALGO == 'native':
+        if SPCONV_ALGO == "native":
             algo = spconv.ConvAlgo.Native
-        elif SPCONV_ALGO == 'implicit_gemm':
+        elif SPCONV_ALGO == "implicit_gemm":
             algo = spconv.ConvAlgo.MaskImplicitGemm
         if stride == 1 and (padding is None):
-            self.conv = spconv.SubMConv3d(in_channels, out_channels, kernel_size, dilation=dilation, bias=bias, indice_key=indice_key, algo=algo)
+            self.conv = spconv.SubMConv3d(
+                in_channels,
+                out_channels,
+                kernel_size,
+                dilation=dilation,
+                bias=bias,
+                indice_key=indice_key,
+                algo=algo,
+            )
         else:
-            self.conv = spconv.SparseConv3d(in_channels, out_channels, kernel_size, stride=stride, dilation=dilation, padding=padding, bias=bias, indice_key=indice_key, algo=algo)
-        self.stride = tuple(stride) if isinstance(stride, (list, tuple)) else (stride, stride, stride)
+            self.conv = spconv.SparseConv3d(
+                in_channels,
+                out_channels,
+                kernel_size,
+                stride=stride,
+                dilation=dilation,
+                padding=padding,
+                bias=bias,
+                indice_key=indice_key,
+                algo=algo,
+            )
+        self.stride = (
+            tuple(stride)
+            if isinstance(stride, (list, tuple))
+            else (stride, stride, stride)
+        )
         self.padding = padding
 
     def forward(self, x: SparseTensor) -> SparseTensor:
@@ -30,42 +63,65 @@ class SparseConv3d(nn.Module):
         if spatial_changed and (x.shape[0] != 1):
             # spconv was non-1 stride will break the contiguous of the output tensor, sort by the coords
             fwd = new_data.indices[:, 0].argsort()
-            bwd = torch.zeros_like(fwd).scatter_(0, fwd, torch.arange(fwd.shape[0], device=fwd.device))
+            bwd = torch.zeros_like(fwd).scatter_(
+                0, fwd, torch.arange(fwd.shape[0], device=fwd.device)
+            )
             sorted_feats = new_data.features[fwd]
             sorted_coords = new_data.indices[fwd]
             unsorted_data = new_data
             new_data = spconv.SparseConvTensor(sorted_feats, sorted_coords, unsorted_data.spatial_shape, unsorted_data.batch_size)  # type: ignore
 
         out = SparseTensor(
-            new_data, shape=torch.Size(new_shape), layout=new_layout,
+            new_data,
+            shape=torch.Size(new_shape),
+            layout=new_layout,
             scale=tuple([s * stride for s, stride in zip(x._scale, self.stride)]),
             spatial_cache=x._spatial_cache,
         )
 
         if spatial_changed and (x.shape[0] != 1):
-            out.register_spatial_cache(f'conv_{self.stride}_unsorted_data', unsorted_data)
-            out.register_spatial_cache(f'conv_{self.stride}_sort_bwd', bwd)
- 
+            out.register_spatial_cache(
+                f"conv_{self.stride}_unsorted_data", unsorted_data
+            )
+            out.register_spatial_cache(f"conv_{self.stride}_sort_bwd", bwd)
+
         return out
 
 
 class SparseInverseConv3d(nn.Module):
-    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        dilation=1,
+        bias=True,
+        indice_key=None,
+    ):
         super(SparseInverseConv3d, self).__init__()
-        if 'spconv' not in globals():
+        if "spconv" not in globals():
             import spconv.pytorch as spconv
-        self.conv = spconv.SparseInverseConv3d(in_channels, out_channels, kernel_size, bias=bias, indice_key=indice_key)
-        self.stride = tuple(stride) if isinstance(stride, (list, tuple)) else (stride, stride, stride)
+        self.conv = spconv.SparseInverseConv3d(
+            in_channels, out_channels, kernel_size, bias=bias, indice_key=indice_key
+        )
+        self.stride = (
+            tuple(stride)
+            if isinstance(stride, (list, tuple))
+            else (stride, stride, stride)
+        )
 
     def forward(self, x: SparseTensor) -> SparseTensor:
         spatial_changed = any(s != 1 for s in self.stride)
         if spatial_changed:
             # recover the original spconv order
-            data = x.get_spatial_cache(f'conv_{self.stride}_unsorted_data')
-            bwd = x.get_spatial_cache(f'conv_{self.stride}_sort_bwd')
+            data = x.get_spatial_cache(f"conv_{self.stride}_unsorted_data")
+            bwd = x.get_spatial_cache(f"conv_{self.stride}_sort_bwd")
             data = data.replace_feature(x.feats[bwd])
             if DEBUG:
-                assert torch.equal(data.indices, x.coords[bwd]), 'Recover the original order failed'
+                assert torch.equal(
+                    data.indices, x.coords[bwd]
+                ), "Recover the original order failed"
         else:
             data = x.data
 
@@ -73,7 +129,9 @@ class SparseInverseConv3d(nn.Module):
         new_shape = [x.shape[0], self.conv.out_channels]
         new_layout = None if spatial_changed else x.layout
         out = SparseTensor(
-            new_data, shape=torch.Size(new_shape), layout=new_layout,
+            new_data,
+            shape=torch.Size(new_shape),
+            layout=new_layout,
             scale=tuple([s // stride for s, stride in zip(x._scale, self.stride)]),
             spatial_cache=x._spatial_cache,
         )

trellis/modules/sparse/conv/conv_torchsparse.py

@@ -4,35 +4,73 @@ from .. import SparseTensor
 
 
 class SparseConv3d(nn.Module):
-    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        dilation=1,
+        bias=True,
+        indice_key=None,
+    ):
         super(SparseConv3d, self).__init__()
-        if 'torchsparse' not in globals():
+        if "torchsparse" not in globals():
             import torchsparse
-        self.conv = torchsparse.nn.Conv3d(in_channels, out_channels, kernel_size, stride, 0, dilation, bias)
+        self.conv = torchsparse.nn.Conv3d(
+            in_channels, out_channels, kernel_size, stride, 0, dilation, bias
+        )
 
     def forward(self, x: SparseTensor) -> SparseTensor:
         out = self.conv(x.data)
         new_shape = [x.shape[0], self.conv.out_channels]
-        out = SparseTensor(out, shape=torch.Size(new_shape), layout=x.layout if all(s == 1 for s in self.conv.stride) else None)
+        out = SparseTensor(
+            out,
+            shape=torch.Size(new_shape),
+            layout=x.layout if all(s == 1 for s in self.conv.stride) else None,
+        )
         out._spatial_cache = x._spatial_cache
-        out._scale = tuple([s * stride for s, stride in zip(x._scale, self.conv.stride)])
+        out._scale = tuple(
+            [s * stride for s, stride in zip(x._scale, self.conv.stride)]
+        )
         return out
 
 
 class SparseInverseConv3d(nn.Module):
-    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        dilation=1,
+        bias=True,
+        indice_key=None,
+    ):
         super(SparseInverseConv3d, self).__init__()
-        if 'torchsparse' not in globals():
+        if "torchsparse" not in globals():
             import torchsparse
-        self.conv = torchsparse.nn.Conv3d(in_channels, out_channels, kernel_size, stride, 0, dilation, bias, transposed=True)
+        self.conv = torchsparse.nn.Conv3d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            0,
+            dilation,
+            bias,
+            transposed=True,
+        )
 
     def forward(self, x: SparseTensor) -> SparseTensor:
-        out = self.conv(x.data)        
+        out = self.conv(x.data)
         new_shape = [x.shape[0], self.conv.out_channels]
-        out = SparseTensor(out, shape=torch.Size(new_shape), layout=x.layout if all(s == 1 for s in self.conv.stride) else None)
+        out = SparseTensor(
+            out,
+            shape=torch.Size(new_shape),
+            layout=x.layout if all(s == 1 for s in self.conv.stride) else None,
+        )
         out._spatial_cache = x._spatial_cache
-        out._scale = tuple([s // stride for s, stride in zip(x._scale, self.conv.stride)])
+        out._scale = tuple(
+            [s // stride for s, stride in zip(x._scale, self.conv.stride)]
+        )
         return out
-
-
-

trellis/modules/sparse/linear.py

@@ -2,9 +2,7 @@ import torch
 import torch.nn as nn
 from . import SparseTensor
 
-__all__ = [
-    'SparseLinear'
-]
+__all__ = ["SparseLinear"]
 
 
 class SparseLinear(nn.Linear):

trellis/modules/sparse/nonlinearity.py

@@ -2,18 +2,13 @@ import torch
 import torch.nn as nn
 from . import SparseTensor
 
-__all__ = [
-    'SparseReLU',
-    'SparseSiLU',
-    'SparseGELU',
-    'SparseActivation'
-]
+__all__ = ["SparseReLU", "SparseSiLU", "SparseGELU", "SparseActivation"]
 
 
 class SparseReLU(nn.ReLU):
     def forward(self, input: SparseTensor) -> SparseTensor:
         return input.replace(super().forward(input.feats))
-    
+
 
 class SparseSiLU(nn.SiLU):
     def forward(self, input: SparseTensor) -> SparseTensor:
@@ -32,4 +27,3 @@ class SparseActivation(nn.Module):
 
     def forward(self, input: SparseTensor) -> SparseTensor:
         return input.replace(self.activation(input.feats))
-    

trellis/modules/sparse/norm.py

@@ -4,10 +4,10 @@ from . import SparseTensor
 from . import DEBUG
 
 __all__ = [
-    'SparseGroupNorm',
-    'SparseLayerNorm',
-    'SparseGroupNorm32',
-    'SparseLayerNorm32',
+    "SparseGroupNorm",
+    "SparseLayerNorm",
+    "SparseGroupNorm32",
+    "SparseLayerNorm32",
 ]
 
 
@@ -19,7 +19,9 @@ class SparseGroupNorm(nn.GroupNorm):
         nfeats = torch.zeros_like(input.feats)
         for k in range(input.shape[0]):
             if DEBUG:
-                assert (input.coords[input.layout[k], 0] == k).all(), f"SparseGroupNorm: batch index mismatch"
+                assert (
+                    input.coords[input.layout[k], 0] == k
+                ).all(), f"SparseGroupNorm: batch index mismatch"
             bfeats = input.feats[input.layout[k]]
             bfeats = bfeats.permute(1, 0).reshape(1, input.shape[1], -1)
             bfeats = super().forward(bfeats)
@@ -47,12 +49,15 @@ class SparseGroupNorm32(SparseGroupNorm):
     """
     A GroupNorm layer that converts to float32 before the forward pass.
     """
+
     def forward(self, x: SparseTensor) -> SparseTensor:
         return super().forward(x.float()).type(x.dtype)
 
+
 class SparseLayerNorm32(SparseLayerNorm):
     """
     A LayerNorm layer that converts to float32 before the forward pass.
     """
+
     def forward(self, x: SparseTensor) -> SparseTensor:
         return super().forward(x.float()).type(x.dtype)

trellis/modules/sparse/spatial.py

@@ -3,11 +3,7 @@ import torch
 import torch.nn as nn
 from . import SparseTensor
 
-__all__ = [
-    'SparseDownsample',
-    'SparseUpsample',
-    'SparseSubdivide'
-]
+__all__ = ["SparseDownsample", "SparseUpsample", "SparseSubdivide"]
 
 
 class SparseDownsample(nn.Module):
@@ -15,6 +11,7 @@ class SparseDownsample(nn.Module):
     Downsample a sparse tensor by a factor of `factor`.
     Implemented as average pooling.
     """
+
     def __init__(self, factor: Union[int, Tuple[int, ...], List[int]]):
         super(SparseDownsample, self).__init__()
         self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
@@ -22,36 +19,47 @@ class SparseDownsample(nn.Module):
     def forward(self, input: SparseTensor) -> SparseTensor:
         DIM = input.coords.shape[-1] - 1
         factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
-        assert DIM == len(factor), 'Input coordinates must have the same dimension as the downsample factor.'
+        assert DIM == len(
+            factor
+        ), "Input coordinates must have the same dimension as the downsample factor."
 
         coord = list(input.coords.unbind(dim=-1))
         for i, f in enumerate(factor):
-            coord[i+1] = coord[i+1] // f
+            coord[i + 1] = coord[i + 1] // f
 
-        MAX = [coord[i+1].max().item() + 1 for i in range(DIM)]
+        MAX = [coord[i + 1].max().item() + 1 for i in range(DIM)]
         OFFSET = torch.cumprod(torch.tensor(MAX[::-1]), 0).tolist()[::-1] + [1]
         code = sum([c * o for c, o in zip(coord, OFFSET)])
         code, idx = code.unique(return_inverse=True)
 
         new_feats = torch.scatter_reduce(
-            torch.zeros(code.shape[0], input.feats.shape[1], device=input.feats.device, dtype=input.feats.dtype),
+            torch.zeros(
+                code.shape[0],
+                input.feats.shape[1],
+                device=input.feats.device,
+                dtype=input.feats.dtype,
+            ),
             dim=0,
             index=idx.unsqueeze(1).expand(-1, input.feats.shape[1]),
             src=input.feats,
-            reduce='mean'
+            reduce="mean",
         )
         new_coords = torch.stack(
-            [code // OFFSET[0]] +
-            [(code // OFFSET[i+1]) % MAX[i] for i in range(DIM)],
-            dim=-1
+            [code // OFFSET[0]]
+            + [(code // OFFSET[i + 1]) % MAX[i] for i in range(DIM)],
+            dim=-1,
+        )
+        out = SparseTensor(
+            new_feats,
+            new_coords,
+            input.shape,
         )
-        out = SparseTensor(new_feats, new_coords, input.shape,)
         out._scale = tuple([s // f for s, f in zip(input._scale, factor)])
         out._spatial_cache = input._spatial_cache
 
-        out.register_spatial_cache(f'upsample_{factor}_coords', input.coords)
-        out.register_spatial_cache(f'upsample_{factor}_layout', input.layout)
-        out.register_spatial_cache(f'upsample_{factor}_idx', idx)
+        out.register_spatial_cache(f"upsample_{factor}_coords", input.coords)
+        out.register_spatial_cache(f"upsample_{factor}_layout", input.layout)
+        out.register_spatial_cache(f"upsample_{factor}_idx", idx)
 
         return out
 
@@ -61,6 +69,7 @@ class SparseUpsample(nn.Module):
     Upsample a sparse tensor by a factor of `factor`.
     Implemented as nearest neighbor interpolation.
     """
+
     def __init__(self, factor: Union[int, Tuple[int, int, int], List[int]]):
         super(SparseUpsample, self).__init__()
         self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
@@ -68,24 +77,30 @@ class SparseUpsample(nn.Module):
     def forward(self, input: SparseTensor) -> SparseTensor:
         DIM = input.coords.shape[-1] - 1
         factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
-        assert DIM == len(factor), 'Input coordinates must have the same dimension as the upsample factor.'
+        assert DIM == len(
+            factor
+        ), "Input coordinates must have the same dimension as the upsample factor."
 
-        new_coords = input.get_spatial_cache(f'upsample_{factor}_coords')
-        new_layout = input.get_spatial_cache(f'upsample_{factor}_layout')
-        idx = input.get_spatial_cache(f'upsample_{factor}_idx')
+        new_coords = input.get_spatial_cache(f"upsample_{factor}_coords")
+        new_layout = input.get_spatial_cache(f"upsample_{factor}_layout")
+        idx = input.get_spatial_cache(f"upsample_{factor}_idx")
         if any([x is None for x in [new_coords, new_layout, idx]]):
-            raise ValueError('Upsample cache not found. SparseUpsample must be paired with SparseDownsample.')
+            raise ValueError(
+                "Upsample cache not found. SparseUpsample must be paired with SparseDownsample."
+            )
         new_feats = input.feats[idx]
         out = SparseTensor(new_feats, new_coords, input.shape, new_layout)
         out._scale = tuple([s * f for s, f in zip(input._scale, factor)])
         out._spatial_cache = input._spatial_cache
         return out
-    
+
+
 class SparseSubdivide(nn.Module):
     """
     Upsample a sparse tensor by a factor of `factor`.
     Implemented as nearest neighbor interpolation.
     """
+
     def __init__(self):
         super(SparseSubdivide, self).__init__()
 
@@ -96,15 +111,20 @@ class SparseSubdivide(nn.Module):
         n_coords = torch.nonzero(n_cube)
         n_coords = torch.cat([torch.zeros_like(n_coords[:, :1]), n_coords], dim=-1)
         factor = n_coords.shape[0]
-        assert factor == 2 ** DIM
+        assert factor == 2**DIM
         # print(n_coords.shape)
         new_coords = input.coords.clone()
         new_coords[:, 1:] *= 2
-        new_coords = new_coords.unsqueeze(1) + n_coords.unsqueeze(0).to(new_coords.dtype)
-        
-        new_feats = input.feats.unsqueeze(1).expand(input.feats.shape[0], factor, *input.feats.shape[1:])
-        out = SparseTensor(new_feats.flatten(0, 1), new_coords.flatten(0, 1), input.shape)
+        new_coords = new_coords.unsqueeze(1) + n_coords.unsqueeze(0).to(
+            new_coords.dtype
+        )
+
+        new_feats = input.feats.unsqueeze(1).expand(
+            input.feats.shape[0], factor, *input.feats.shape[1:]
+        )
+        out = SparseTensor(
+            new_feats.flatten(0, 1), new_coords.flatten(0, 1), input.shape
+        )
         out._scale = input._scale * 2
         out._spatial_cache = input._spatial_cache
         return out
-

trellis/modules/sparse/transformer/__init__.py

@@ -1,2 +1,2 @@
 from .blocks import *
-from .modulated import *
+from .modulated import *

trellis/modules/sparse/transformer/blocks.py

@@ -25,12 +25,15 @@ class SparseTransformerBlock(nn.Module):
     """
     Sparse Transformer block (MSA + FFN).
     """
+
     def __init__(
         self,
         channels: int,
         num_heads: int,
         mlp_ratio: float = 4.0,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "full",
         window_size: Optional[int] = None,
         shift_sequence: Optional[int] = None,
         shift_window: Optional[Tuple[int, int, int]] = None,
@@ -73,7 +76,9 @@ class SparseTransformerBlock(nn.Module):
 
     def forward(self, x: SparseTensor) -> SparseTensor:
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, use_reentrant=False
+            )
         else:
             return self._forward(x)
 
@@ -82,13 +87,16 @@ class SparseTransformerCrossBlock(nn.Module):
     """
     Sparse Transformer cross-attention block (MSA + MCA + FFN).
     """
+
     def __init__(
         self,
         channels: int,
         ctx_channels: int,
         num_heads: int,
         mlp_ratio: float = 4.0,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "full",
         window_size: Optional[int] = None,
         shift_sequence: Optional[int] = None,
         shift_window: Optional[Tuple[int, int, int]] = None,
@@ -146,6 +154,8 @@ class SparseTransformerCrossBlock(nn.Module):
 
     def forward(self, x: SparseTensor, context: torch.Tensor):
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, context, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, context, use_reentrant=False
+            )
         else:
             return self._forward(x, context)

trellis/modules/sparse/transformer/modulated.py

@@ -11,12 +11,15 @@ class ModulatedSparseTransformerBlock(nn.Module):
     """
     Sparse Transformer block (MSA + FFN) with adaptive layer norm conditioning.
     """
+
     def __init__(
         self,
         channels: int,
         num_heads: int,
         mlp_ratio: float = 4.0,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "full",
         window_size: Optional[int] = None,
         shift_sequence: Optional[int] = None,
         shift_window: Optional[Tuple[int, int, int]] = None,
@@ -50,15 +53,18 @@ class ModulatedSparseTransformerBlock(nn.Module):
         )
         if not share_mod:
             self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                nn.Linear(channels, 6 * channels, bias=True)
+                nn.SiLU(), nn.Linear(channels, 6 * channels, bias=True)
             )
 
     def _forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
         if self.share_mod:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
+                6, dim=1
+            )
         else:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.adaLN_modulation(mod).chunk(6, dim=1)
+            )
         h = x.replace(self.norm1(x.feats))
         h = h * (1 + scale_msa) + shift_msa
         h = self.attn(h)
@@ -73,7 +79,9 @@ class ModulatedSparseTransformerBlock(nn.Module):
 
     def forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, mod, use_reentrant=False
+            )
         else:
             return self._forward(x, mod)
 
@@ -82,13 +90,16 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
     """
     Sparse Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
     """
+
     def __init__(
         self,
         channels: int,
         ctx_channels: int,
         num_heads: int,
         mlp_ratio: float = 4.0,
-        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        attn_mode: Literal[
+            "full", "shift_window", "shift_sequence", "shift_order", "swin"
+        ] = "full",
         window_size: Optional[int] = None,
         shift_sequence: Optional[int] = None,
         shift_window: Optional[Tuple[int, int, int]] = None,
@@ -99,7 +110,6 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
         qk_rms_norm_cross: bool = False,
         qkv_bias: bool = True,
         share_mod: bool = False,
-
     ):
         super().__init__()
         self.use_checkpoint = use_checkpoint
@@ -135,15 +145,20 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
         )
         if not share_mod:
             self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                nn.Linear(channels, 6 * channels, bias=True)
+                nn.SiLU(), nn.Linear(channels, 6 * channels, bias=True)
             )
 
-    def _forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor) -> SparseTensor:
+    def _forward(
+        self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor
+    ) -> SparseTensor:
         if self.share_mod:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
+                6, dim=1
+            )
         else:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.adaLN_modulation(mod).chunk(6, dim=1)
+            )
         h = x.replace(self.norm1(x.feats))
         h = h * (1 + scale_msa) + shift_msa
         h = self.self_attn(h)
@@ -159,8 +174,12 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
         x = x + h
         return x
 
-    def forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor) -> SparseTensor:
+    def forward(
+        self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor
+    ) -> SparseTensor:
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, context, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, mod, context, use_reentrant=False
+            )
         else:
             return self._forward(x, mod, context)

trellis/modules/spatial.py

@@ -9,7 +9,7 @@ def pixel_shuffle_3d(x: torch.Tensor, scale_factor: int) -> torch.Tensor:
     C_ = C // scale_factor**3
     x = x.reshape(B, C_, scale_factor, scale_factor, scale_factor, H, W, D)
     x = x.permute(0, 1, 5, 2, 6, 3, 7, 4)
-    x = x.reshape(B, C_, H*scale_factor, W*scale_factor, D*scale_factor)
+    x = x.reshape(B, C_, H * scale_factor, W * scale_factor, D * scale_factor)
     return x
 
 
@@ -23,11 +23,18 @@ def patchify(x: torch.Tensor, patch_size: int):
     """
     DIM = x.dim() - 2
     for d in range(2, DIM + 2):
-        assert x.shape[d] % patch_size == 0, f"Dimension {d} of input tensor must be divisible by patch size, got {x.shape[d]} and {patch_size}"
+        assert (
+            x.shape[d] % patch_size == 0
+        ), f"Dimension {d} of input tensor must be divisible by patch size, got {x.shape[d]} and {patch_size}"
 
-    x = x.reshape(*x.shape[:2], *sum([[x.shape[d] // patch_size, patch_size] for d in range(2, DIM + 2)], []))
-    x = x.permute(0, 1, *([2 * i + 3 for i in range(DIM)] + [2 * i + 2 for i in range(DIM)]))
-    x = x.reshape(x.shape[0], x.shape[1] * (patch_size ** DIM), *(x.shape[-DIM:]))
+    x = x.reshape(
+        *x.shape[:2],
+        *sum([[x.shape[d] // patch_size, patch_size] for d in range(2, DIM + 2)], []),
+    )
+    x = x.permute(
+        0, 1, *([2 * i + 3 for i in range(DIM)] + [2 * i + 2 for i in range(DIM)])
+    )
+    x = x.reshape(x.shape[0], x.shape[1] * (patch_size**DIM), *(x.shape[-DIM:]))
     return x
 
 
@@ -40,9 +47,18 @@ def unpatchify(x: torch.Tensor, patch_size: int):
         patch_size (int): Patch size
     """
     DIM = x.dim() - 2
-    assert x.shape[1] % (patch_size ** DIM) == 0, f"Second dimension of input tensor must be divisible by patch size to unpatchify, got {x.shape[1]} and {patch_size ** DIM}"
+    assert (
+        x.shape[1] % (patch_size**DIM) == 0
+    ), f"Second dimension of input tensor must be divisible by patch size to unpatchify, got {x.shape[1]} and {patch_size ** DIM}"
 
-    x = x.reshape(x.shape[0], x.shape[1] // (patch_size ** DIM), *([patch_size] * DIM), *(x.shape[-DIM:]))
+    x = x.reshape(
+        x.shape[0],
+        x.shape[1] // (patch_size**DIM),
+        *([patch_size] * DIM),
+        *(x.shape[-DIM:]),
+    )
     x = x.permute(0, 1, *(sum([[2 + DIM + i, 2 + i] for i in range(DIM)], [])))
-    x = x.reshape(x.shape[0], x.shape[1], *[x.shape[2 + 2 * i] * patch_size for i in range(DIM)])
+    x = x.reshape(
+        x.shape[0], x.shape[1], *[x.shape[2 + 2 * i] * patch_size for i in range(DIM)]
+    )
     return x

trellis/modules/transformer/__init__.py

@@ -1,2 +1,2 @@
 from .blocks import *
-from .modulated import *
+from .modulated import *

trellis/modules/transformer/blocks.py

@@ -9,14 +9,15 @@ class AbsolutePositionEmbedder(nn.Module):
     """
     Embeds spatial positions into vector representations.
     """
+
     def __init__(self, channels: int, in_channels: int = 3):
         super().__init__()
         self.channels = channels
         self.in_channels = in_channels
         self.freq_dim = channels // in_channels // 2
         self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
-        self.freqs = 1.0 / (10000 ** self.freqs)
-        
+        self.freqs = 1.0 / (10000**self.freqs)
+
     def _sin_cos_embedding(self, x: torch.Tensor) -> torch.Tensor:
         """
         Create sinusoidal position embeddings.
@@ -38,11 +39,19 @@ class AbsolutePositionEmbedder(nn.Module):
             x (torch.Tensor): (N, D) tensor of spatial positions
         """
         N, D = x.shape
-        assert D == self.in_channels, "Input dimension must match number of input channels"
+        assert (
+            D == self.in_channels
+        ), "Input dimension must match number of input channels"
         embed = self._sin_cos_embedding(x.reshape(-1))
         embed = embed.reshape(N, -1)
         if embed.shape[1] < self.channels:
-            embed = torch.cat([embed, torch.zeros(N, self.channels - embed.shape[1], device=embed.device)], dim=-1)
+            embed = torch.cat(
+                [
+                    embed,
+                    torch.zeros(N, self.channels - embed.shape[1], device=embed.device),
+                ],
+                dim=-1,
+            )
         return embed
 
 
@@ -63,6 +72,7 @@ class TransformerBlock(nn.Module):
     """
     Transformer block (MSA + FFN).
     """
+
     def __init__(
         self,
         channels: int,
@@ -107,7 +117,9 @@ class TransformerBlock(nn.Module):
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, use_reentrant=False
+            )
         else:
             return self._forward(x)
 
@@ -116,6 +128,7 @@ class TransformerCrossBlock(nn.Module):
     """
     Transformer cross-attention block (MSA + MCA + FFN).
     """
+
     def __init__(
         self,
         channels: int,
@@ -176,7 +189,8 @@ class TransformerCrossBlock(nn.Module):
 
     def forward(self, x: torch.Tensor, context: torch.Tensor):
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, context, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, context, use_reentrant=False
+            )
         else:
             return self._forward(x, context)
-        

trellis/modules/transformer/modulated.py

@@ -10,6 +10,7 @@ class ModulatedTransformerBlock(nn.Module):
     """
     Transformer block (MSA + FFN) with adaptive layer norm conditioning.
     """
+
     def __init__(
         self,
         channels: int,
@@ -45,15 +46,18 @@ class ModulatedTransformerBlock(nn.Module):
         )
         if not share_mod:
             self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                nn.Linear(channels, 6 * channels, bias=True)
+                nn.SiLU(), nn.Linear(channels, 6 * channels, bias=True)
             )
 
     def _forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:
         if self.share_mod:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
+                6, dim=1
+            )
         else:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.adaLN_modulation(mod).chunk(6, dim=1)
+            )
         h = self.norm1(x)
         h = h * (1 + scale_msa.unsqueeze(1)) + shift_msa.unsqueeze(1)
         h = self.attn(h)
@@ -68,7 +72,9 @@ class ModulatedTransformerBlock(nn.Module):
 
     def forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, mod, use_reentrant=False
+            )
         else:
             return self._forward(x, mod)
 
@@ -77,6 +83,7 @@ class ModulatedTransformerCrossBlock(nn.Module):
     """
     Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
     """
+
     def __init__(
         self,
         channels: int,
@@ -125,15 +132,18 @@ class ModulatedTransformerCrossBlock(nn.Module):
         )
         if not share_mod:
             self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                nn.Linear(channels, 6 * channels, bias=True)
+                nn.SiLU(), nn.Linear(channels, 6 * channels, bias=True)
             )
 
     def _forward(self, x: torch.Tensor, mod: torch.Tensor, context: torch.Tensor):
         if self.share_mod:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
+                6, dim=1
+            )
         else:
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.adaLN_modulation(mod).chunk(6, dim=1)
+            )
         h = self.norm1(x)
         h = h * (1 + scale_msa.unsqueeze(1)) + shift_msa.unsqueeze(1)
         h = self.self_attn(h)
@@ -151,7 +161,8 @@ class ModulatedTransformerCrossBlock(nn.Module):
 
     def forward(self, x: torch.Tensor, mod: torch.Tensor, context: torch.Tensor):
         if self.use_checkpoint:
-            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, context, use_reentrant=False)
+            return torch.utils.checkpoint.checkpoint(
+                self._forward, x, mod, context, use_reentrant=False
+            )
         else:
             return self._forward(x, mod, context)
-        

trellis/modules/utils.py

@@ -14,6 +14,7 @@ FP16_MODULES = (
     sp.SparseLinear,
 )
 
+
 def convert_module_to_f16(l):
     """
     Convert primitive modules to float16.

trellis/pipelines/__init__.py

@@ -11,14 +11,16 @@ def from_pretrained(path: str):
     """
     import os
     import json
+
     is_local = os.path.exists(f"{path}/pipeline.json")
 
     if is_local:
         config_file = f"{path}/pipeline.json"
     else:
         from huggingface_hub import hf_hub_download
+
         config_file = hf_hub_download(path, "pipeline.json")
 
-    with open(config_file, 'r') as f:
+    with open(config_file, "r") as f:
         config = json.load(f)
-    return globals()[config['name']].from_pretrained(path)
+    return globals()[config["name"]].from_pretrained(path)

trellis/pipelines/base.py

@@ -8,6 +8,7 @@ class Pipeline:
     """
     A base class for pipelines.
     """
+
     def __init__(
         self,
         models: dict[str, nn.Module] = None,
@@ -25,20 +26,21 @@ class Pipeline:
         """
         import os
         import json
+
         is_local = os.path.exists(f"{path}/pipeline.json")
 
         if is_local:
             config_file = f"{path}/pipeline.json"
         else:
             from huggingface_hub import hf_hub_download
+
             config_file = hf_hub_download(path, "pipeline.json")
 
-        with open(config_file, 'r') as f:
-            args = json.load(f)['args']
+        with open(config_file, "r") as f:
+            args = json.load(f)["args"]
 
         _models = {
-            k: models.from_pretrained(f"{path}/{v}")
-            for k, v in args['models'].items()
+            k: models.from_pretrained(f"{path}/{v}") for k, v in args["models"].items()
         }
 
         new_pipeline = Pipeline(_models)
@@ -48,10 +50,10 @@ class Pipeline:
     @property
     def device(self) -> torch.device:
         for model in self.models.values():
-            if hasattr(model, 'device'):
+            if hasattr(model, "device"):
                 return model.device
         for model in self.models.values():
-            if hasattr(model, 'parameters'):
+            if hasattr(model, "parameters"):
                 return next(model.parameters()).device
         raise RuntimeError("No device found.")
 

trellis/pipelines/samplers/__init__.py

@@ -1,2 +1,6 @@
 from .base import Sampler
-from .flow_euler import FlowEulerSampler, FlowEulerCfgSampler, FlowEulerGuidanceIntervalSampler
+from .flow_euler import (
+    FlowEulerSampler,
+    FlowEulerCfgSampler,
+    FlowEulerGuidanceIntervalSampler,
+)

trellis/pipelines/samplers/base.py

@@ -8,13 +8,8 @@ class Sampler(ABC):
     """
 
     @abstractmethod
-    def sample(
-        self,
-        model,
-        **kwargs
-    ):
+    def sample(self, model, **kwargs):
         """
         Sample from a model.
         """
         pass
-    

trellis/pipelines/samplers/flow_euler.py

@@ -15,6 +15,7 @@ class FlowEulerSampler(Sampler):
     Args:
         sigma_min: The minimum scale of noise in flow.
     """
+
     def __init__(
         self,
         sigma_min: float,
@@ -32,11 +33,15 @@ class FlowEulerSampler(Sampler):
     def _v_to_xstart_eps(self, x_t, t, v):
         assert x_t.shape == v.shape
         eps = (1 - t) * v + x_t
-        x_0 = (1 - self.sigma_min) * x_t - (self.sigma_min + (1 - self.sigma_min) * t) * v
+        x_0 = (1 - self.sigma_min) * x_t - (
+            self.sigma_min + (1 - self.sigma_min) * t
+        ) * v
         return x_0, eps
 
     def _inference_model(self, model, x_t, t, cond=None, **kwargs):
-        t = torch.tensor([1000 * t] * x_t.shape[0], device=x_t.device, dtype=torch.float32)
+        t = torch.tensor(
+            [1000 * t] * x_t.shape[0], device=x_t.device, dtype=torch.float32
+        )
         return model(x_t, t, cond, **kwargs)
 
     def _get_model_prediction(self, model, x_t, t, cond=None, **kwargs):
@@ -46,17 +51,11 @@ class FlowEulerSampler(Sampler):
 
     @torch.no_grad()
     def sample_once(
-        self,
-        model,
-        x_t,
-        t: float,
-        t_prev: float,
-        cond: Optional[Any] = None,
-        **kwargs
+        self, model, x_t, t: float, t_prev: float, cond: Optional[Any] = None, **kwargs
     ):
         """
         Sample x_{t-1} from the model using Euler method.
-        
+
         Args:
             model: The model to sample from.
             x_t: The [N x C x ...] tensor of noisy inputs at time t.
@@ -70,7 +69,9 @@ class FlowEulerSampler(Sampler):
             - 'pred_x_prev': x_{t-1}.
             - 'pred_x_0': a prediction of x_0.
         """
-        pred_x_0, pred_eps, pred_v = self._get_model_prediction(model, x_t, t, cond, **kwargs)
+        pred_x_0, pred_eps, pred_v = self._get_model_prediction(
+            model, x_t, t, cond, **kwargs
+        )
         pred_x_prev = x_t - (t - t_prev) * pred_v
         return edict({"pred_x_prev": pred_x_prev, "pred_x_0": pred_x_0})
 
@@ -87,7 +88,7 @@ class FlowEulerSampler(Sampler):
     ):
         """
         Generate samples from the model using Euler method.
-        
+
         Args:
             model: The model to sample from.
             noise: The initial noise tensor.
@@ -121,6 +122,7 @@ class FlowEulerCfgSampler(ClassifierFreeGuidanceSamplerMixin, FlowEulerSampler):
     """
     Generate samples from a flow-matching model using Euler sampling with classifier-free guidance.
     """
+
     @torch.no_grad()
     def sample(
         self,
@@ -136,7 +138,7 @@ class FlowEulerCfgSampler(ClassifierFreeGuidanceSamplerMixin, FlowEulerSampler):
     ):
         """
         Generate samples from the model using Euler method.
-        
+
         Args:
             model: The model to sample from.
             noise: The initial noise tensor.
@@ -154,13 +156,24 @@ class FlowEulerCfgSampler(ClassifierFreeGuidanceSamplerMixin, FlowEulerSampler):
             - 'pred_x_t': a list of prediction of x_t.
             - 'pred_x_0': a list of prediction of x_0.
         """
-        return super().sample(model, noise, cond, steps, rescale_t, verbose, neg_cond=neg_cond, cfg_strength=cfg_strength, **kwargs)
+        return super().sample(
+            model,
+            noise,
+            cond,
+            steps,
+            rescale_t,
+            verbose,
+            neg_cond=neg_cond,
+            cfg_strength=cfg_strength,
+            **kwargs
+        )
 
 
 class FlowEulerGuidanceIntervalSampler(GuidanceIntervalSamplerMixin, FlowEulerSampler):
     """
     Generate samples from a flow-matching model using Euler sampling with classifier-free guidance and interval.
     """
+
     @torch.no_grad()
     def sample(
         self,
@@ -177,7 +190,7 @@ class FlowEulerGuidanceIntervalSampler(GuidanceIntervalSamplerMixin, FlowEulerSa
     ):
         """
         Generate samples from the model using Euler method.
-        
+
         Args:
             model: The model to sample from.
             noise: The initial noise tensor.
@@ -196,4 +209,15 @@ class FlowEulerGuidanceIntervalSampler(GuidanceIntervalSamplerMixin, FlowEulerSa
             - 'pred_x_t': a list of prediction of x_t.
             - 'pred_x_0': a list of prediction of x_0.
         """
-        return super().sample(model, noise, cond, steps, rescale_t, verbose, neg_cond=neg_cond, cfg_strength=cfg_strength, cfg_interval=cfg_interval, **kwargs)
+        return super().sample(
+            model,
+            noise,
+            cond,
+            steps,
+            rescale_t,
+            verbose,
+            neg_cond=neg_cond,
+            cfg_strength=cfg_strength,
+            cfg_interval=cfg_interval,
+            **kwargs
+        )

trellis/pipelines/samplers/guidance_interval_mixin.py

@@ -6,7 +6,9 @@ class GuidanceIntervalSamplerMixin:
     A mixin class for samplers that apply classifier-free guidance with interval.
     """
 
-    def _inference_model(self, model, x_t, t, cond, neg_cond, cfg_strength, cfg_interval, **kwargs):
+    def _inference_model(
+        self, model, x_t, t, cond, neg_cond, cfg_strength, cfg_interval, **kwargs
+    ):
         if cfg_interval[0] <= t <= cfg_interval[1]:
             pred = super()._inference_model(model, x_t, t, cond, **kwargs)
             neg_pred = super()._inference_model(model, x_t, t, neg_cond, **kwargs)

trellis/pipelines/trellis_image_to_3d.py

@@ -26,6 +26,7 @@ class TrellisImageTo3DPipeline(Pipeline):
         slat_normalization (dict): The normalization parameters for the structured latent.
         image_cond_model (str): The name of the image conditioning model.
     """
+
     def __init__(
         self,
         models: dict[str, nn.Module] = None,
@@ -53,33 +54,45 @@ class TrellisImageTo3DPipeline(Pipeline):
         Args:
             path (str): The path to the model. Can be either local path or a Hugging Face repository.
         """
-        pipeline = super(TrellisImageTo3DPipeline, TrellisImageTo3DPipeline).from_pretrained(path)
+        pipeline = super(
+            TrellisImageTo3DPipeline, TrellisImageTo3DPipeline
+        ).from_pretrained(path)
         new_pipeline = TrellisImageTo3DPipeline()
         new_pipeline.__dict__ = pipeline.__dict__
         args = pipeline._pretrained_args
 
-        new_pipeline.sparse_structure_sampler = getattr(samplers, args['sparse_structure_sampler']['name'])(**args['sparse_structure_sampler']['args'])
-        new_pipeline.sparse_structure_sampler_params = args['sparse_structure_sampler']['params']
+        new_pipeline.sparse_structure_sampler = getattr(
+            samplers, args["sparse_structure_sampler"]["name"]
+        )(**args["sparse_structure_sampler"]["args"])
+        new_pipeline.sparse_structure_sampler_params = args["sparse_structure_sampler"][
+            "params"
+        ]
 
-        new_pipeline.slat_sampler = getattr(samplers, args['slat_sampler']['name'])(**args['slat_sampler']['args'])
-        new_pipeline.slat_sampler_params = args['slat_sampler']['params']
+        new_pipeline.slat_sampler = getattr(samplers, args["slat_sampler"]["name"])(
+            **args["slat_sampler"]["args"]
+        )
+        new_pipeline.slat_sampler_params = args["slat_sampler"]["params"]
 
-        new_pipeline.slat_normalization = args['slat_normalization']
+        new_pipeline.slat_normalization = args["slat_normalization"]
 
-        new_pipeline._init_image_cond_model(args['image_cond_model'])
+        new_pipeline._init_image_cond_model(args["image_cond_model"])
 
         return new_pipeline
-    
+
     def _init_image_cond_model(self, name: str):
         """
         Initialize the image conditioning model.
         """
-        dinov2_model = torch.hub.load('facebookresearch/dinov2', name, pretrained=True)
+        dinov2_model = torch.hub.load("facebookresearch/dinov2", name, pretrained=True)
         dinov2_model.eval()
-        self.models['image_cond_model'] = dinov2_model
-        transform = transforms.Compose([
-            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-        ])
+        self.models["image_cond_model"] = dinov2_model
+        transform = transforms.Compose(
+            [
+                transforms.Normalize(
+                    mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+                ),
+            ]
+        )
         self.image_cond_model_transform = transform
 
     def preprocess_image(self, input: Image.Image) -> Image.Image:
@@ -88,29 +101,42 @@ class TrellisImageTo3DPipeline(Pipeline):
         """
         # if has alpha channel, use it directly; otherwise, remove background
         has_alpha = False
-        if input.mode == 'RGBA':
+        if input.mode == "RGBA":
             alpha = np.array(input)[:, :, 3]
             if not np.all(alpha == 255):
                 has_alpha = True
         if has_alpha:
             output = input
         else:
-            input = input.convert('RGB')
+            input = input.convert("RGB")
             max_size = max(input.size)
             scale = min(1, 1024 / max_size)
             if scale < 1:
-                input = input.resize((int(input.width * scale), int(input.height * scale)), Image.Resampling.LANCZOS)
-            if getattr(self, 'rembg_session', None) is None:
-                self.rembg_session = rembg.new_session('u2net')
+                input = input.resize(
+                    (int(input.width * scale), int(input.height * scale)),
+                    Image.Resampling.LANCZOS,
+                )
+            if getattr(self, "rembg_session", None) is None:
+                self.rembg_session = rembg.new_session("u2net")
             output = rembg.remove(input, session=self.rembg_session)
         output_np = np.array(output)
         alpha = output_np[:, :, 3]
         bbox = np.argwhere(alpha > 0.8 * 255)
-        bbox = np.min(bbox[:, 1]), np.min(bbox[:, 0]), np.max(bbox[:, 1]), np.max(bbox[:, 0])
+        bbox = (
+            np.min(bbox[:, 1]),
+            np.min(bbox[:, 0]),
+            np.max(bbox[:, 1]),
+            np.max(bbox[:, 0]),
+        )
         center = (bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2
         size = max(bbox[2] - bbox[0], bbox[3] - bbox[1])
         size = int(size * 1.2)
-        bbox = center[0] - size // 2, center[1] - size // 2, center[0] + size // 2, center[1] + size // 2
+        bbox = (
+            center[0] - size // 2,
+            center[1] - size // 2,
+            center[0] + size // 2,
+            center[1] + size // 2,
+        )
         output = output.crop(bbox)  # type: ignore
         output = output.resize((518, 518), Image.Resampling.LANCZOS)
         output = np.array(output).astype(np.float32) / 255
@@ -119,7 +145,9 @@ class TrellisImageTo3DPipeline(Pipeline):
         return output
 
     @torch.no_grad()
-    def encode_image(self, image: Union[torch.Tensor, list[Image.Image]]) -> torch.Tensor:
+    def encode_image(
+        self, image: Union[torch.Tensor, list[Image.Image]]
+    ) -> torch.Tensor:
         """
         Encode the image.
 
@@ -132,19 +160,21 @@ class TrellisImageTo3DPipeline(Pipeline):
         if isinstance(image, torch.Tensor):
             assert image.ndim == 4, "Image tensor should be batched (B, C, H, W)"
         elif isinstance(image, list):
-            assert all(isinstance(i, Image.Image) for i in image), "Image list should be list of PIL images"
+            assert all(
+                isinstance(i, Image.Image) for i in image
+            ), "Image list should be list of PIL images"
             image = [i.resize((518, 518), Image.LANCZOS) for i in image]
-            image = [np.array(i.convert('RGB')).astype(np.float32) / 255 for i in image]
+            image = [np.array(i.convert("RGB")).astype(np.float32) / 255 for i in image]
             image = [torch.from_numpy(i).permute(2, 0, 1).float() for i in image]
             image = torch.stack(image).to(self.device)
         else:
             raise ValueError(f"Unsupported type of image: {type(image)}")
-        
+
         image = self.image_cond_model_transform(image).to(self.device)
-        features = self.models['image_cond_model'](image, is_training=True)['x_prenorm']
+        features = self.models["image_cond_model"](image, is_training=True)["x_prenorm"]
         patchtokens = F.layer_norm(features, features.shape[-1:])
         return patchtokens
-        
+
     def get_cond(self, image: Union[torch.Tensor, list[Image.Image]]) -> dict:
         """
         Get the conditioning information for the model.
@@ -158,8 +188,8 @@ class TrellisImageTo3DPipeline(Pipeline):
         cond = self.encode_image(image)
         neg_cond = torch.zeros_like(cond)
         return {
-            'cond': cond,
-            'neg_cond': neg_cond,
+            "cond": cond,
+            "neg_cond": neg_cond,
         }
 
     def sample_sparse_structure(
@@ -170,35 +200,33 @@ class TrellisImageTo3DPipeline(Pipeline):
     ) -> torch.Tensor:
         """
         Sample sparse structures with the given conditioning.
-        
+
         Args:
             cond (dict): The conditioning information.
             num_samples (int): The number of samples to generate.
             sampler_params (dict): Additional parameters for the sampler.
         """
         # Sample occupancy latent
-        flow_model = self.models['sparse_structure_flow_model']
+        flow_model = self.models["sparse_structure_flow_model"]
         reso = flow_model.resolution
-        noise = torch.randn(num_samples, flow_model.in_channels, reso, reso, reso).to(self.device)
+        noise = torch.randn(num_samples, flow_model.in_channels, reso, reso, reso).to(
+            self.device
+        )
         sampler_params = {**self.sparse_structure_sampler_params, **sampler_params}
         z_s = self.sparse_structure_sampler.sample(
-            flow_model,
-            noise,
-            **cond,
-            **sampler_params,
-            verbose=True
+            flow_model, noise, **cond, **sampler_params, verbose=True
         ).samples
-        
+
         # Decode occupancy latent
-        decoder = self.models['sparse_structure_decoder']
-        coords = torch.argwhere(decoder(z_s)>0)[:, [0, 2, 3, 4]].int()
+        decoder = self.models["sparse_structure_decoder"]
+        coords = torch.argwhere(decoder(z_s) > 0)[:, [0, 2, 3, 4]].int()
 
         return coords
 
     def decode_slat(
         self,
         slat: sp.SparseTensor,
-        formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
+        formats: List[str] = ["mesh", "gaussian", "radiance_field"],
     ) -> dict:
         """
         Decode the structured latent.
@@ -211,14 +239,14 @@ class TrellisImageTo3DPipeline(Pipeline):
             dict: The decoded structured latent.
         """
         ret = {}
-        if 'mesh' in formats:
-            ret['mesh'] = self.models['slat_decoder_mesh'](slat)
-        if 'gaussian' in formats:
-            ret['gaussian'] = self.models['slat_decoder_gs'](slat)
-        if 'radiance_field' in formats:
-            ret['radiance_field'] = self.models['slat_decoder_rf'](slat)
+        if "mesh" in formats:
+            ret["mesh"] = self.models["slat_decoder_mesh"](slat)
+        if "gaussian" in formats:
+            ret["gaussian"] = self.models["slat_decoder_gs"](slat)
+        if "radiance_field" in formats:
+            ret["radiance_field"] = self.models["slat_decoder_rf"](slat)
         return ret
-    
+
     def sample_slat(
         self,
         cond: dict,
@@ -227,31 +255,27 @@ class TrellisImageTo3DPipeline(Pipeline):
     ) -> sp.SparseTensor:
         """
         Sample structured latent with the given conditioning.
-        
+
         Args:
             cond (dict): The conditioning information.
             coords (torch.Tensor): The coordinates of the sparse structure.
             sampler_params (dict): Additional parameters for the sampler.
         """
         # Sample structured latent
-        flow_model = self.models['slat_flow_model']
+        flow_model = self.models["slat_flow_model"]
         noise = sp.SparseTensor(
             feats=torch.randn(coords.shape[0], flow_model.in_channels).to(self.device),
             coords=coords,
         )
         sampler_params = {**self.slat_sampler_params, **sampler_params}
         slat = self.slat_sampler.sample(
-            flow_model,
-            noise,
-            **cond,
-            **sampler_params,
-            verbose=True
+            flow_model, noise, **cond, **sampler_params, verbose=True
         ).samples
 
-        std = torch.tensor(self.slat_normalization['std'])[None].to(slat.device)
-        mean = torch.tensor(self.slat_normalization['mean'])[None].to(slat.device)
+        std = torch.tensor(self.slat_normalization["std"])[None].to(slat.device)
+        mean = torch.tensor(self.slat_normalization["mean"])[None].to(slat.device)
         slat = slat * std + mean
-        
+
         return slat
 
     @torch.no_grad()
@@ -262,7 +286,7 @@ class TrellisImageTo3DPipeline(Pipeline):
         seed: int = 42,
         sparse_structure_sampler_params: dict = {},
         slat_sampler_params: dict = {},
-        formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
+        formats: List[str] = ["mesh", "gaussian", "radiance_field"],
         preprocess_image: bool = True,
     ) -> dict:
         """
@@ -279,7 +303,9 @@ class TrellisImageTo3DPipeline(Pipeline):
             image = self.preprocess_image(image)
         cond = self.get_cond([image])
         torch.manual_seed(seed)
-        coords = self.sample_sparse_structure(cond, num_samples, sparse_structure_sampler_params)
+        coords = self.sample_sparse_structure(
+            cond, num_samples, sparse_structure_sampler_params
+        )
         slat = self.sample_slat(cond, coords, slat_sampler_params)
         return self.decode_slat(slat, formats)
 
@@ -289,56 +315,80 @@ class TrellisImageTo3DPipeline(Pipeline):
         sampler_name: str,
         num_images: int,
         num_steps: int,
-        mode: Literal['stochastic', 'multidiffusion'] = 'stochastic',
+        mode: Literal["stochastic", "multidiffusion"] = "stochastic",
     ):
         """
         Inject a sampler with multiple images as condition.
-        
+
         Args:
             sampler_name (str): The name of the sampler to inject.
             num_images (int): The number of images to condition on.
             num_steps (int): The number of steps to run the sampler for.
         """
         sampler = getattr(self, sampler_name)
-        setattr(sampler, f'_old_inference_model', sampler._inference_model)
+        setattr(sampler, f"_old_inference_model", sampler._inference_model)
 
-        if mode == 'stochastic':
+        if mode == "stochastic":
             if num_images > num_steps:
-                print(f"\033[93mWarning: number of conditioning images is greater than number of steps for {sampler_name}. "
-                    "This may lead to performance degradation.\033[0m")
+                print(
+                    f"\033[93mWarning: number of conditioning images is greater than number of steps for {sampler_name}. "
+                    "This may lead to performance degradation.\033[0m"
+                )
 
             cond_indices = (np.arange(num_steps) % num_images).tolist()
+
             def _new_inference_model(self, model, x_t, t, cond, **kwargs):
                 cond_idx = cond_indices.pop(0)
-                cond_i = cond[cond_idx:cond_idx+1]
+                cond_i = cond[cond_idx : cond_idx + 1]
                 return self._old_inference_model(model, x_t, t, cond=cond_i, **kwargs)
-        
-        elif mode =='multidiffusion':
+
+        elif mode == "multidiffusion":
             from .samplers import FlowEulerSampler
-            def _new_inference_model(self, model, x_t, t, cond, neg_cond, cfg_strength, cfg_interval, **kwargs):
+
+            def _new_inference_model(
+                self,
+                model,
+                x_t,
+                t,
+                cond,
+                neg_cond,
+                cfg_strength,
+                cfg_interval,
+                **kwargs,
+            ):
                 if cfg_interval[0] <= t <= cfg_interval[1]:
                     preds = []
                     for i in range(len(cond)):
-                        preds.append(FlowEulerSampler._inference_model(self, model, x_t, t, cond[i:i+1], **kwargs))
+                        preds.append(
+                            FlowEulerSampler._inference_model(
+                                self, model, x_t, t, cond[i : i + 1], **kwargs
+                            )
+                        )
                     pred = sum(preds) / len(preds)
-                    neg_pred = FlowEulerSampler._inference_model(self, model, x_t, t, neg_cond, **kwargs)
+                    neg_pred = FlowEulerSampler._inference_model(
+                        self, model, x_t, t, neg_cond, **kwargs
+                    )
                     return (1 + cfg_strength) * pred - cfg_strength * neg_pred
                 else:
                     preds = []
                     for i in range(len(cond)):
-                        preds.append(FlowEulerSampler._inference_model(self, model, x_t, t, cond[i:i+1], **kwargs))
+                        preds.append(
+                            FlowEulerSampler._inference_model(
+                                self, model, x_t, t, cond[i : i + 1], **kwargs
+                            )
+                        )
                     pred = sum(preds) / len(preds)
                     return pred
-            
+
         else:
             raise ValueError(f"Unsupported mode: {mode}")
-            
+
         sampler._inference_model = _new_inference_model.__get__(sampler, type(sampler))
 
         yield
 
         sampler._inference_model = sampler._old_inference_model
-        delattr(sampler, f'_old_inference_model')
+        delattr(sampler, f"_old_inference_model")
 
     @torch.no_grad()
     def run_multi_image(
@@ -348,9 +398,9 @@ class TrellisImageTo3DPipeline(Pipeline):
         seed: int = 42,
         sparse_structure_sampler_params: dict = {},
         slat_sampler_params: dict = {},
-        formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
+        formats: List[str] = ["mesh", "gaussian", "radiance_field"],
         preprocess_image: bool = True,
-        mode: Literal['stochastic', 'multidiffusion'] = 'stochastic',
+        mode: Literal["stochastic", "multidiffusion"] = "stochastic",
     ) -> dict:
         """
         Run the pipeline with multiple images as condition
@@ -365,12 +415,21 @@ class TrellisImageTo3DPipeline(Pipeline):
         if preprocess_image:
             images = [self.preprocess_image(image) for image in images]
         cond = self.get_cond(images)
-        cond['neg_cond'] = cond['neg_cond'][:1]
+        cond["neg_cond"] = cond["neg_cond"][:1]
         torch.manual_seed(seed)
-        ss_steps = {**self.sparse_structure_sampler_params, **sparse_structure_sampler_params}.get('steps')
-        with self.inject_sampler_multi_image('sparse_structure_sampler', len(images), ss_steps, mode=mode):
-            coords = self.sample_sparse_structure(cond, num_samples, sparse_structure_sampler_params)
-        slat_steps = {**self.slat_sampler_params, **slat_sampler_params}.get('steps')
-        with self.inject_sampler_multi_image('slat_sampler', len(images), slat_steps, mode=mode):
+        ss_steps = {
+            **self.sparse_structure_sampler_params,
+            **sparse_structure_sampler_params,
+        }.get("steps")
+        with self.inject_sampler_multi_image(
+            "sparse_structure_sampler", len(images), ss_steps, mode=mode
+        ):
+            coords = self.sample_sparse_structure(
+                cond, num_samples, sparse_structure_sampler_params
+            )
+        slat_steps = {**self.slat_sampler_params, **slat_sampler_params}.get("steps")
+        with self.inject_sampler_multi_image(
+            "slat_sampler", len(images), slat_steps, mode=mode
+        ):
             slat = self.sample_slat(cond, coords, slat_sampler_params)
         return self.decode_slat(slat, formats)

trellis/renderers/__init__.py

@@ -1,15 +1,16 @@
 import importlib
 
 __attributes = {
-    'OctreeRenderer': 'octree_renderer',
-    'GaussianRenderer': 'gaussian_render',
-    'MeshRenderer': 'mesh_renderer',
+    "OctreeRenderer": "octree_renderer",
+    "GaussianRenderer": "gaussian_render",
+    "MeshRenderer": "mesh_renderer",
 }
 
 __submodules = []
 
 __all__ = list(__attributes.keys()) + __submodules
 
+
 def __getattr__(name):
     if name not in globals():
         if name in __attributes:
@@ -25,7 +26,7 @@ def __getattr__(name):
 
 
 # For Pylance
-if __name__ == '__main__':
+if __name__ == "__main__":
     from .octree_renderer import OctreeRenderer
     from .gaussian_render import GaussianRenderer
-    from .mesh_renderer import MeshRenderer
+    from .mesh_renderer import MeshRenderer

trellis/renderers/gaussian_render.py

@@ -3,7 +3,7 @@
 # GRAPHDECO research group, https://team.inria.fr/graphdeco
 # All rights reserved.
 #
-# This software is free for non-commercial, research and evaluation use 
+# This software is free for non-commercial, research and evaluation use
 # under the terms of the LICENSE.md file.
 #
 # For inquiries contact  [email protected]
@@ -20,10 +20,10 @@ from easydict import EasyDict as edict
 
 
 def intrinsics_to_projection(
-        intrinsics: torch.Tensor,
-        near: float,
-        far: float,
-    ) -> torch.Tensor:
+    intrinsics: torch.Tensor,
+    near: float,
+    far: float,
+) -> torch.Tensor:
     """
     OpenCV intrinsics to OpenGL perspective matrix
 
@@ -40,25 +40,40 @@ def intrinsics_to_projection(
     ret[0, 0] = 2 * fx
     ret[1, 1] = 2 * fy
     ret[0, 2] = 2 * cx - 1
-    ret[1, 2] = - 2 * cy + 1
+    ret[1, 2] = -2 * cy + 1
     ret[2, 2] = far / (far - near)
     ret[2, 3] = near * far / (near - far)
-    ret[3, 2] = 1.
+    ret[3, 2] = 1.0
     return ret
 
 
-def render(viewpoint_camera, pc : Gaussian, pipe, bg_color : torch.Tensor, scaling_modifier = 1.0, override_color = None):
+def render(
+    viewpoint_camera,
+    pc: Gaussian,
+    pipe,
+    bg_color: torch.Tensor,
+    scaling_modifier=1.0,
+    override_color=None,
+):
     """
-    Render the scene. 
-    
+    Render the scene.
+
     Background tensor (bg_color) must be on GPU!
     """
     # lazy import
-    if 'GaussianRasterizer' not in globals():
-        from diff_gaussian_rasterization import GaussianRasterizer, GaussianRasterizationSettings
-    
+    if "GaussianRasterizer" not in globals():
+        from diff_gaussian_rasterization import (
+            GaussianRasterizer,
+            GaussianRasterizationSettings,
+        )
+
     # Create zero tensor. We will use it to make pytorch return gradients of the 2D (screen-space) means
-    screenspace_points = torch.zeros_like(pc.get_xyz, dtype=pc.get_xyz.dtype, requires_grad=True, device="cuda") + 0
+    screenspace_points = (
+        torch.zeros_like(
+            pc.get_xyz, dtype=pc.get_xyz.dtype, requires_grad=True, device="cuda"
+        )
+        + 0
+    )
     try:
         screenspace_points.retain_grad()
     except:
@@ -66,9 +81,13 @@ def render(viewpoint_camera, pc : Gaussian, pipe, bg_color : torch.Tensor, scali
     # Set up rasterization configuration
     tanfovx = math.tan(viewpoint_camera.FoVx * 0.5)
     tanfovy = math.tan(viewpoint_camera.FoVy * 0.5)
-    
+
     kernel_size = pipe.kernel_size
-    subpixel_offset = torch.zeros((int(viewpoint_camera.image_height), int(viewpoint_camera.image_width), 2), dtype=torch.float32, device="cuda")
+    subpixel_offset = torch.zeros(
+        (int(viewpoint_camera.image_height), int(viewpoint_camera.image_width), 2),
+        dtype=torch.float32,
+        device="cuda",
+    )
 
     raster_settings = GaussianRasterizationSettings(
         image_height=int(viewpoint_camera.image_height),
@@ -84,9 +103,9 @@ def render(viewpoint_camera, pc : Gaussian, pipe, bg_color : torch.Tensor, scali
         sh_degree=pc.active_sh_degree,
         campos=viewpoint_camera.camera_center,
         prefiltered=False,
-        debug=pipe.debug
+        debug=pipe.debug,
     )
-    
+
     rasterizer = GaussianRasterizer(raster_settings=raster_settings)
 
     means3D = pc.get_xyz
@@ -110,9 +129,13 @@ def render(viewpoint_camera, pc : Gaussian, pipe, bg_color : torch.Tensor, scali
     colors_precomp = None
     if override_color is None:
         if pipe.convert_SHs_python:
-            shs_view = pc.get_features.transpose(1, 2).view(-1, 3, (pc.max_sh_degree+1)**2)
-            dir_pp = (pc.get_xyz - viewpoint_camera.camera_center.repeat(pc.get_features.shape[0], 1))
-            dir_pp_normalized = dir_pp/dir_pp.norm(dim=1, keepdim=True)
+            shs_view = pc.get_features.transpose(1, 2).view(
+                -1, 3, (pc.max_sh_degree + 1) ** 2
+            )
+            dir_pp = pc.get_xyz - viewpoint_camera.camera_center.repeat(
+                pc.get_features.shape[0], 1
+            )
+            dir_pp_normalized = dir_pp / dir_pp.norm(dim=1, keepdim=True)
             sh2rgb = eval_sh(pc.active_sh_degree, shs_view, dir_pp_normalized)
             colors_precomp = torch.clamp_min(sh2rgb + 0.5, 0.0)
         else:
@@ -120,24 +143,28 @@ def render(viewpoint_camera, pc : Gaussian, pipe, bg_color : torch.Tensor, scali
     else:
         colors_precomp = override_color
 
-    # Rasterize visible Gaussians to image, obtain their radii (on screen). 
+    # Rasterize visible Gaussians to image, obtain their radii (on screen).
     rendered_image, radii = rasterizer(
-        means3D = means3D,
-        means2D = means2D,
-        shs = shs,
-        colors_precomp = colors_precomp,
-        opacities = opacity,
-        scales = scales,
-        rotations = rotations,
-        cov3D_precomp = cov3D_precomp
+        means3D=means3D,
+        means2D=means2D,
+        shs=shs,
+        colors_precomp=colors_precomp,
+        opacities=opacity,
+        scales=scales,
+        rotations=rotations,
+        cov3D_precomp=cov3D_precomp,
     )
 
     # Those Gaussians that were frustum culled or had a radius of 0 were not visible.
     # They will be excluded from value updates used in the splitting criteria.
-    return edict({"render": rendered_image,
+    return edict(
+        {
+            "render": rendered_image,
             "viewspace_points": screenspace_points,
-            "visibility_filter" : radii > 0,
-            "radii": radii})
+            "visibility_filter": radii > 0,
+            "radii": radii,
+        }
+    )
 
 
 class GaussianRenderer:
@@ -149,30 +176,34 @@ class GaussianRenderer:
     """
 
     def __init__(self, rendering_options={}) -> None:
-        self.pipe = edict({
-            "kernel_size": 0.1,
-            "convert_SHs_python": False,
-            "compute_cov3D_python": False,
-            "scale_modifier": 1.0,
-            "debug": False
-        })
-        self.rendering_options = edict({
-            "resolution": None,
-            "near": None,
-            "far": None,
-            "ssaa": 1,
-            "bg_color": 'random',
-        })
+        self.pipe = edict(
+            {
+                "kernel_size": 0.1,
+                "convert_SHs_python": False,
+                "compute_cov3D_python": False,
+                "scale_modifier": 1.0,
+                "debug": False,
+            }
+        )
+        self.rendering_options = edict(
+            {
+                "resolution": None,
+                "near": None,
+                "far": None,
+                "ssaa": 1,
+                "bg_color": "random",
+            }
+        )
         self.rendering_options.update(rendering_options)
         self.bg_color = None
-    
+
     def render(
-            self,
-            gausssian: Gaussian,
-            extrinsics: torch.Tensor,
-            intrinsics: torch.Tensor,
-            colors_overwrite: torch.Tensor = None
-        ) -> edict:
+        self,
+        gausssian: Gaussian,
+        extrinsics: torch.Tensor,
+        intrinsics: torch.Tensor,
+        colors_overwrite: torch.Tensor = None,
+    ) -> edict:
         """
         Render the gausssian.
 
@@ -190,13 +221,15 @@ class GaussianRenderer:
         near = self.rendering_options["near"]
         far = self.rendering_options["far"]
         ssaa = self.rendering_options["ssaa"]
-        
-        if self.rendering_options["bg_color"] == 'random':
+
+        if self.rendering_options["bg_color"] == "random":
             self.bg_color = torch.zeros(3, dtype=torch.float32, device="cuda")
             if np.random.rand() < 0.5:
                 self.bg_color += 1
         else:
-            self.bg_color = torch.tensor(self.rendering_options["bg_color"], dtype=torch.float32, device="cuda")
+            self.bg_color = torch.tensor(
+                self.rendering_options["bg_color"], dtype=torch.float32, device="cuda"
+            )
 
         view = extrinsics
         perspective = intrinsics_to_projection(intrinsics, near, far)
@@ -205,27 +238,40 @@ class GaussianRenderer:
         focaly = intrinsics[1, 1]
         fovx = 2 * torch.atan(0.5 / focalx)
         fovy = 2 * torch.atan(0.5 / focaly)
-            
-        camera_dict = edict({
-            "image_height": resolution * ssaa,
-            "image_width": resolution * ssaa,
-            "FoVx": fovx,
-            "FoVy": fovy,
-            "znear": near,
-            "zfar": far,
-            "world_view_transform": view.T.contiguous(),
-            "projection_matrix": perspective.T.contiguous(),
-            "full_proj_transform": (perspective @ view).T.contiguous(),
-            "camera_center": camera
-        })
+
+        camera_dict = edict(
+            {
+                "image_height": resolution * ssaa,
+                "image_width": resolution * ssaa,
+                "FoVx": fovx,
+                "FoVy": fovy,
+                "znear": near,
+                "zfar": far,
+                "world_view_transform": view.T.contiguous(),
+                "projection_matrix": perspective.T.contiguous(),
+                "full_proj_transform": (perspective @ view).T.contiguous(),
+                "camera_center": camera,
+            }
+        )
 
         # Render
-        render_ret = render(camera_dict, gausssian, self.pipe, self.bg_color, override_color=colors_overwrite, scaling_modifier=self.pipe.scale_modifier)
+        render_ret = render(
+            camera_dict,
+            gausssian,
+            self.pipe,
+            self.bg_color,
+            override_color=colors_overwrite,
+            scaling_modifier=self.pipe.scale_modifier,
+        )
 
         if ssaa > 1:
-            render_ret.render = F.interpolate(render_ret.render[None], size=(resolution, resolution), mode='bilinear', align_corners=False, antialias=True).squeeze()
-            
-        ret = edict({
-            'color': render_ret['render']
-        })
+            render_ret.render = F.interpolate(
+                render_ret.render[None],
+                size=(resolution, resolution),
+                mode="bilinear",
+                align_corners=False,
+                antialias=True,
+            ).squeeze()
+
+        ret = edict({"color": render_ret["render"]})
         return ret

trellis/renderers/mesh_renderer.py

@@ -13,10 +13,10 @@ import torch.nn.functional as F
 
 
 def intrinsics_to_projection(
-        intrinsics: torch.Tensor,
-        near: float,
-        far: float,
-    ) -> torch.Tensor:
+    intrinsics: torch.Tensor,
+    near: float,
+    far: float,
+) -> torch.Tensor:
     """
     OpenCV intrinsics to OpenGL perspective matrix
 
@@ -33,10 +33,10 @@ def intrinsics_to_projection(
     ret[0, 0] = 2 * fx
     ret[1, 1] = 2 * fy
     ret[0, 2] = 2 * cx - 1
-    ret[1, 2] = - 2 * cy + 1
+    ret[1, 2] = -2 * cy + 1
     ret[2, 2] = far / (far - near)
     ret[2, 3] = near * far / (near - far)
-    ret[3, 2] = 1.
+    ret[3, 2] = 1.0
     return ret
 
 
@@ -47,25 +47,23 @@ class MeshRenderer:
     Args:
         rendering_options (dict): Rendering options.
         glctx (nvdiffrast.torch.RasterizeGLContext): RasterizeGLContext object for CUDA/OpenGL interop.
-        """
-    def __init__(self, rendering_options={}, device='cuda'):
-        self.rendering_options = edict({
-            "resolution": None,
-            "near": None,
-            "far": None,
-            "ssaa": 1
-        })
+    """
+
+    def __init__(self, rendering_options={}, device="cuda"):
+        self.rendering_options = edict(
+            {"resolution": None, "near": None, "far": None, "ssaa": 1}
+        )
         self.rendering_options.update(rendering_options)
         self.glctx = dr.RasterizeCudaContext(device=device)
-        self.device=device
-        
+        self.device = device
+
     def render(
-            self,
-            mesh : MeshExtractResult,
-            extrinsics: torch.Tensor,
-            intrinsics: torch.Tensor,
-            return_types = ["mask", "normal", "depth"]
-        ) -> edict:
+        self,
+        mesh: MeshExtractResult,
+        extrinsics: torch.Tensor,
+        intrinsics: torch.Tensor,
+        return_types=["mask", "normal", "depth"],
+    ) -> edict:
         """
         Render the mesh.
 
@@ -87,51 +85,80 @@ class MeshRenderer:
         near = self.rendering_options["near"]
         far = self.rendering_options["far"]
         ssaa = self.rendering_options["ssaa"]
-        
+
         if mesh.vertices.shape[0] == 0 or mesh.faces.shape[0] == 0:
-            default_img = torch.zeros((1, resolution, resolution, 3), dtype=torch.float32, device=self.device)
-            ret_dict = {k : default_img if k in ['normal', 'normal_map', 'color'] else default_img[..., :1] for k in return_types}
+            default_img = torch.zeros(
+                (1, resolution, resolution, 3), dtype=torch.float32, device=self.device
+            )
+            ret_dict = {
+                k: (
+                    default_img
+                    if k in ["normal", "normal_map", "color"]
+                    else default_img[..., :1]
+                )
+                for k in return_types
+            }
             return ret_dict
-        
+
         perspective = intrinsics_to_projection(intrinsics, near, far)
-        
+
         RT = extrinsics.unsqueeze(0)
         full_proj = (perspective @ extrinsics).unsqueeze(0)
-        
+
         vertices = mesh.vertices.unsqueeze(0)
 
-        vertices_homo = torch.cat([vertices, torch.ones_like(vertices[..., :1])], dim=-1)
+        vertices_homo = torch.cat(
+            [vertices, torch.ones_like(vertices[..., :1])], dim=-1
+        )
         vertices_camera = torch.bmm(vertices_homo, RT.transpose(-1, -2))
         vertices_clip = torch.bmm(vertices_homo, full_proj.transpose(-1, -2))
         faces_int = mesh.faces.int()
         rast, _ = dr.rasterize(
-            self.glctx, vertices_clip, faces_int, (resolution * ssaa, resolution * ssaa))
-        
+            self.glctx, vertices_clip, faces_int, (resolution * ssaa, resolution * ssaa)
+        )
+
         out_dict = edict()
         for type in return_types:
             img = None
-            if type == "mask" :
-                img = dr.antialias((rast[..., -1:] > 0).float(), rast, vertices_clip, faces_int)
+            if type == "mask":
+                img = dr.antialias(
+                    (rast[..., -1:] > 0).float(), rast, vertices_clip, faces_int
+                )
             elif type == "depth":
-                img = dr.interpolate(vertices_camera[..., 2:3].contiguous(), rast, faces_int)[0]
+                img = dr.interpolate(
+                    vertices_camera[..., 2:3].contiguous(), rast, faces_int
+                )[0]
                 img = dr.antialias(img, rast, vertices_clip, faces_int)
-            elif type == "normal" :
+            elif type == "normal":
                 img = dr.interpolate(
-                    mesh.face_normal.reshape(1, -1, 3), rast,
-                    torch.arange(mesh.faces.shape[0] * 3, device=self.device, dtype=torch.int).reshape(-1, 3)
+                    mesh.face_normal.reshape(1, -1, 3),
+                    rast,
+                    torch.arange(
+                        mesh.faces.shape[0] * 3, device=self.device, dtype=torch.int
+                    ).reshape(-1, 3),
                 )[0]
                 img = dr.antialias(img, rast, vertices_clip, faces_int)
                 # normalize norm pictures
                 img = (img + 1) / 2
-            elif type == "normal_map" :
-                img = dr.interpolate(mesh.vertex_attrs[:, 3:].contiguous(), rast, faces_int)[0]
+            elif type == "normal_map":
+                img = dr.interpolate(
+                    mesh.vertex_attrs[:, 3:].contiguous(), rast, faces_int
+                )[0]
                 img = dr.antialias(img, rast, vertices_clip, faces_int)
-            elif type == "color" :
-                img = dr.interpolate(mesh.vertex_attrs[:, :3].contiguous(), rast, faces_int)[0]
+            elif type == "color":
+                img = dr.interpolate(
+                    mesh.vertex_attrs[:, :3].contiguous(), rast, faces_int
+                )[0]
                 img = dr.antialias(img, rast, vertices_clip, faces_int)
 
             if ssaa > 1:
-                img = F.interpolate(img.permute(0, 3, 1, 2), (resolution, resolution), mode='bilinear', align_corners=False, antialias=True)
+                img = F.interpolate(
+                    img.permute(0, 3, 1, 2),
+                    (resolution, resolution),
+                    mode="bilinear",
+                    align_corners=False,
+                    antialias=True,
+                )
                 img = img.squeeze()
             else:
                 img = img.permute(0, 3, 1, 2).squeeze()

trellis/renderers/octree_renderer.py

@@ -9,10 +9,10 @@ from ..representations.octree import DfsOctree
 
 
 def intrinsics_to_projection(
-        intrinsics: torch.Tensor,
-        near: float,
-        far: float,
-    ) -> torch.Tensor:
+    intrinsics: torch.Tensor,
+    near: float,
+    far: float,
+) -> torch.Tensor:
     """
     OpenCV intrinsics to OpenGL perspective matrix
 
@@ -29,23 +29,38 @@ def intrinsics_to_projection(
     ret[0, 0] = 2 * fx
     ret[1, 1] = 2 * fy
     ret[0, 2] = 2 * cx - 1
-    ret[1, 2] = - 2 * cy + 1
+    ret[1, 2] = -2 * cy + 1
     ret[2, 2] = far / (far - near)
     ret[2, 3] = near * far / (near - far)
-    ret[3, 2] = 1.
+    ret[3, 2] = 1.0
     return ret
 
 
-def render(viewpoint_camera, octree : DfsOctree, pipe, bg_color : torch.Tensor, scaling_modifier = 1.0, used_rank = None, colors_overwrite = None, aux=None, halton_sampler=None):
+def render(
+    viewpoint_camera,
+    octree: DfsOctree,
+    pipe,
+    bg_color: torch.Tensor,
+    scaling_modifier=1.0,
+    used_rank=None,
+    colors_overwrite=None,
+    aux=None,
+    halton_sampler=None,
+):
     """
-    Render the scene. 
-    
+    Render the scene.
+
     Background tensor (bg_color) must be on GPU!
     """
     # lazy import
-    if 'OctreeTrivecRasterizer' not in globals():
-        from diffoctreerast import OctreeVoxelRasterizer, OctreeGaussianRasterizer, OctreeTrivecRasterizer, OctreeDecoupolyRasterizer
-    
+    if "OctreeTrivecRasterizer" not in globals():
+        from diffoctreerast import (
+            OctreeVoxelRasterizer,
+            OctreeGaussianRasterizer,
+            OctreeTrivecRasterizer,
+            OctreeDecoupolyRasterizer,
+        )
+
     # Set up rasterization configuration
     tanfovx = math.tan(viewpoint_camera.FoVx * 0.5)
     tanfovy = math.tan(viewpoint_camera.FoVy * 0.5)
@@ -96,69 +111,73 @@ def render(viewpoint_camera, octree : DfsOctree, pipe, bg_color : torch.Tensor,
     if octree.primitive == "voxel":
         renderer = OctreeVoxelRasterizer(raster_settings=raster_settings)
         rgb, depth, alpha, distloss = renderer(
-            positions = positions,
-            densities = densities,
-            shs = shs,
-            colors_precomp = colors_precomp,
-            depths = depths,
-            aabb = octree.aabb,
-            aux = aux,
+            positions=positions,
+            densities=densities,
+            shs=shs,
+            colors_precomp=colors_precomp,
+            depths=depths,
+            aabb=octree.aabb,
+            aux=aux,
         )
-        ret['rgb'] = rgb
-        ret['depth'] = depth
-        ret['alpha'] = alpha
-        ret['distloss'] = distloss
+        ret["rgb"] = rgb
+        ret["depth"] = depth
+        ret["alpha"] = alpha
+        ret["distloss"] = distloss
     elif octree.primitive == "gaussian":
         renderer = OctreeGaussianRasterizer(raster_settings=raster_settings)
         rgb, depth, alpha = renderer(
-            positions = positions,
-            opacities = opacities,
-            shs = shs,
-            colors_precomp = colors_precomp,
-            depths = depths,
-            aabb = octree.aabb,
-            aux = aux,
+            positions=positions,
+            opacities=opacities,
+            shs=shs,
+            colors_precomp=colors_precomp,
+            depths=depths,
+            aabb=octree.aabb,
+            aux=aux,
         )
-        ret['rgb'] = rgb
-        ret['depth'] = depth
-        ret['alpha'] = alpha
+        ret["rgb"] = rgb
+        ret["depth"] = depth
+        ret["alpha"] = alpha
     elif octree.primitive == "trivec":
-        raster_settings.used_rank = used_rank if used_rank is not None else trivecs.shape[1]
+        raster_settings.used_rank = (
+            used_rank if used_rank is not None else trivecs.shape[1]
+        )
         renderer = OctreeTrivecRasterizer(raster_settings=raster_settings)
         rgb, depth, alpha, percent_depth = renderer(
-            positions = positions,
-            trivecs = trivecs,
-            densities = densities,
-            shs = shs,
-            colors_precomp = colors_precomp,
-            colors_overwrite = colors_overwrite,
-            depths = depths,
-            aabb = octree.aabb,
-            aux = aux,
-            halton_sampler = halton_sampler,
+            positions=positions,
+            trivecs=trivecs,
+            densities=densities,
+            shs=shs,
+            colors_precomp=colors_precomp,
+            colors_overwrite=colors_overwrite,
+            depths=depths,
+            aabb=octree.aabb,
+            aux=aux,
+            halton_sampler=halton_sampler,
         )
-        ret['percent_depth'] = percent_depth
-        ret['rgb'] = rgb
-        ret['depth'] = depth
-        ret['alpha'] = alpha
+        ret["percent_depth"] = percent_depth
+        ret["rgb"] = rgb
+        ret["depth"] = depth
+        ret["alpha"] = alpha
     elif octree.primitive == "decoupoly":
-        raster_settings.used_rank = used_rank if used_rank is not None else decoupolys_V.shape[1]
+        raster_settings.used_rank = (
+            used_rank if used_rank is not None else decoupolys_V.shape[1]
+        )
         renderer = OctreeDecoupolyRasterizer(raster_settings=raster_settings)
         rgb, depth, alpha = renderer(
-            positions = positions,
-            decoupolys_V = decoupolys_V,
-            decoupolys_g = decoupolys_g,
-            densities = densities,
-            shs = shs,
-            colors_precomp = colors_precomp,
-            depths = depths,
-            aabb = octree.aabb,
-            aux = aux,
+            positions=positions,
+            decoupolys_V=decoupolys_V,
+            decoupolys_g=decoupolys_g,
+            densities=densities,
+            shs=shs,
+            colors_precomp=colors_precomp,
+            depths=depths,
+            aabb=octree.aabb,
+            aux=aux,
         )
-        ret['rgb'] = rgb
-        ret['depth'] = depth
-        ret['alpha'] = alpha
-    
+        ret["rgb"] = rgb
+        ret["depth"] = depth
+        ret["alpha"] = alpha
+
     return ret
 
 
@@ -174,37 +193,43 @@ class OctreeRenderer:
         try:
             import diffoctreerast
         except ImportError:
-            print("\033[93m[WARNING] diffoctreerast is not installed. The renderer will be disabled.\033[0m")
+            print(
+                "\033[93m[WARNING] diffoctreerast is not installed. The renderer will be disabled.\033[0m"
+            )
             self.unsupported = True
         else:
             self.unsupported = False
-        
-        self.pipe = edict({
-            "with_distloss": False,
-            "with_aux": False,
-            "scale_modifier": 1.0,
-            "used_rank": None,
-            "jitter": False,
-            "debug": False,
-        })
-        self.rendering_options = edict({
-            "resolution": None,
-            "near": None,
-            "far": None,
-            "ssaa": 1,
-            "bg_color": 'random',
-        })
+
+        self.pipe = edict(
+            {
+                "with_distloss": False,
+                "with_aux": False,
+                "scale_modifier": 1.0,
+                "used_rank": None,
+                "jitter": False,
+                "debug": False,
+            }
+        )
+        self.rendering_options = edict(
+            {
+                "resolution": None,
+                "near": None,
+                "far": None,
+                "ssaa": 1,
+                "bg_color": "random",
+            }
+        )
         self.halton_sampler = qmc.Halton(2, scramble=False)
         self.rendering_options.update(rendering_options)
         self.bg_color = None
-    
+
     def render(
-            self,
-            octree: DfsOctree,
-            extrinsics: torch.Tensor,
-            intrinsics: torch.Tensor,
-            colors_overwrite: torch.Tensor = None,
-        ) -> edict:
+        self,
+        octree: DfsOctree,
+        extrinsics: torch.Tensor,
+        intrinsics: torch.Tensor,
+        colors_overwrite: torch.Tensor = None,
+    ) -> edict:
         """
         Render the octree.
 
@@ -227,27 +252,53 @@ class OctreeRenderer:
         near = self.rendering_options["near"]
         far = self.rendering_options["far"]
         ssaa = self.rendering_options["ssaa"]
-        
+
         if self.unsupported:
             image = np.zeros((512, 512, 3), dtype=np.uint8)
-            text_bbox = cv2.getTextSize("Unsupported", cv2.FONT_HERSHEY_SIMPLEX, 2, 3)[0]
+            text_bbox = cv2.getTextSize("Unsupported", cv2.FONT_HERSHEY_SIMPLEX, 2, 3)[
+                0
+            ]
             origin = (512 - text_bbox[0]) // 2, (512 - text_bbox[1]) // 2
-            image = cv2.putText(image, "Unsupported", origin, cv2.FONT_HERSHEY_SIMPLEX, 2, (255, 255, 255), 3, cv2.LINE_AA)
+            image = cv2.putText(
+                image,
+                "Unsupported",
+                origin,
+                cv2.FONT_HERSHEY_SIMPLEX,
+                2,
+                (255, 255, 255),
+                3,
+                cv2.LINE_AA,
+            )
             return {
-                'color': torch.tensor(image, dtype=torch.float32).permute(2, 0, 1) / 255,
+                "color": torch.tensor(image, dtype=torch.float32).permute(2, 0, 1)
+                / 255,
             }
-        
-        if self.rendering_options["bg_color"] == 'random':
+
+        if self.rendering_options["bg_color"] == "random":
             self.bg_color = torch.zeros(3, dtype=torch.float32, device="cuda")
             if np.random.rand() < 0.5:
                 self.bg_color += 1
         else:
-            self.bg_color = torch.tensor(self.rendering_options["bg_color"], dtype=torch.float32, device="cuda")
+            self.bg_color = torch.tensor(
+                self.rendering_options["bg_color"], dtype=torch.float32, device="cuda"
+            )
 
         if self.pipe["with_aux"]:
             aux = {
-                'grad_color2': torch.zeros((octree.num_leaf_nodes, 3), dtype=torch.float32, requires_grad=True, device="cuda") + 0,
-                'contributions': torch.zeros((octree.num_leaf_nodes, 1), dtype=torch.float32, requires_grad=True, device="cuda") + 0,
+                "grad_color2": torch.zeros(
+                    (octree.num_leaf_nodes, 3),
+                    dtype=torch.float32,
+                    requires_grad=True,
+                    device="cuda",
+                )
+                + 0,
+                "contributions": torch.zeros(
+                    (octree.num_leaf_nodes, 1),
+                    dtype=torch.float32,
+                    requires_grad=True,
+                    device="cuda",
+                )
+                + 0,
             }
             for k in aux.keys():
                 aux[k].requires_grad_()
@@ -262,39 +313,77 @@ class OctreeRenderer:
         focaly = intrinsics[1, 1]
         fovx = 2 * torch.atan(0.5 / focalx)
         fovy = 2 * torch.atan(0.5 / focaly)
-            
-        camera_dict = edict({
-            "image_height": resolution * ssaa,
-            "image_width": resolution * ssaa,
-            "FoVx": fovx,
-            "FoVy": fovy,
-            "znear": near,
-            "zfar": far,
-            "world_view_transform": view.T.contiguous(),
-            "projection_matrix": perspective.T.contiguous(),
-            "full_proj_transform": (perspective @ view).T.contiguous(),
-            "camera_center": camera
-        })
+
+        camera_dict = edict(
+            {
+                "image_height": resolution * ssaa,
+                "image_width": resolution * ssaa,
+                "FoVx": fovx,
+                "FoVy": fovy,
+                "znear": near,
+                "zfar": far,
+                "world_view_transform": view.T.contiguous(),
+                "projection_matrix": perspective.T.contiguous(),
+                "full_proj_transform": (perspective @ view).T.contiguous(),
+                "camera_center": camera,
+            }
+        )
 
         # Render
-        render_ret = render(camera_dict, octree, self.pipe, self.bg_color, aux=aux, colors_overwrite=colors_overwrite, scaling_modifier=self.pipe.scale_modifier, used_rank=self.pipe.used_rank, halton_sampler=self.halton_sampler)
+        render_ret = render(
+            camera_dict,
+            octree,
+            self.pipe,
+            self.bg_color,
+            aux=aux,
+            colors_overwrite=colors_overwrite,
+            scaling_modifier=self.pipe.scale_modifier,
+            used_rank=self.pipe.used_rank,
+            halton_sampler=self.halton_sampler,
+        )
 
         if ssaa > 1:
-            render_ret.rgb = F.interpolate(render_ret.rgb[None], size=(resolution, resolution), mode='bilinear', align_corners=False, antialias=True).squeeze()
-            render_ret.depth = F.interpolate(render_ret.depth[None, None], size=(resolution, resolution), mode='bilinear', align_corners=False, antialias=True).squeeze()
-            render_ret.alpha = F.interpolate(render_ret.alpha[None, None], size=(resolution, resolution), mode='bilinear', align_corners=False, antialias=True).squeeze()
-            if hasattr(render_ret, 'percent_depth'):
-                render_ret.percent_depth = F.interpolate(render_ret.percent_depth[None, None], size=(resolution, resolution), mode='bilinear', align_corners=False, antialias=True).squeeze()
+            render_ret.rgb = F.interpolate(
+                render_ret.rgb[None],
+                size=(resolution, resolution),
+                mode="bilinear",
+                align_corners=False,
+                antialias=True,
+            ).squeeze()
+            render_ret.depth = F.interpolate(
+                render_ret.depth[None, None],
+                size=(resolution, resolution),
+                mode="bilinear",
+                align_corners=False,
+                antialias=True,
+            ).squeeze()
+            render_ret.alpha = F.interpolate(
+                render_ret.alpha[None, None],
+                size=(resolution, resolution),
+                mode="bilinear",
+                align_corners=False,
+                antialias=True,
+            ).squeeze()
+            if hasattr(render_ret, "percent_depth"):
+                render_ret.percent_depth = F.interpolate(
+                    render_ret.percent_depth[None, None],
+                    size=(resolution, resolution),
+                    mode="bilinear",
+                    align_corners=False,
+                    antialias=True,
+                ).squeeze()
 
-        ret = edict({
-            'color': render_ret.rgb,
-            'depth': render_ret.depth,
-            'alpha': render_ret.alpha,
-        })
-        if self.pipe["with_distloss"] and 'distloss' in render_ret:
-            ret['distloss'] = render_ret.distloss
+        ret = edict(
+            {
+                "color": render_ret.rgb,
+                "depth": render_ret.depth,
+                "alpha": render_ret.alpha,
+            }
+        )
+        if self.pipe["with_distloss"] and "distloss" in render_ret:
+            ret["distloss"] = render_ret.distloss
         if self.pipe["with_aux"]:
-            ret['aux'] = aux
-        if hasattr(render_ret, 'percent_depth'):
-            ret['percent_depth'] = render_ret.percent_depth
+            ret["aux"] = aux
+        if hasattr(render_ret, "percent_depth"):
+            ret["percent_depth"] = render_ret.percent_depth
         return ret

trellis/renderers/sh_utils.py

@@ -30,7 +30,7 @@ C2 = [
     -1.0925484305920792,
     0.31539156525252005,
     -1.0925484305920792,
-    0.5462742152960396
+    0.5462742152960396,
 ]
 C3 = [
     -0.5900435899266435,
@@ -39,7 +39,7 @@ C3 = [
     0.3731763325901154,
     -0.4570457994644658,
     1.445305721320277,
-    -0.5900435899266435
+    -0.5900435899266435,
 ]
 C4 = [
     2.5033429417967046,
@@ -51,7 +51,7 @@ C4 = [
     0.47308734787878004,
     -1.7701307697799304,
     0.6258357354491761,
-]   
+]
 
 
 def eval_sh(deg, sh, dirs):
@@ -74,45 +74,55 @@ def eval_sh(deg, sh, dirs):
     result = C0 * sh[..., 0]
     if deg > 0:
         x, y, z = dirs[..., 0:1], dirs[..., 1:2], dirs[..., 2:3]
-        result = (result -
-                C1 * y * sh[..., 1] +
-                C1 * z * sh[..., 2] -
-                C1 * x * sh[..., 3])
+        result = (
+            result - C1 * y * sh[..., 1] + C1 * z * sh[..., 2] - C1 * x * sh[..., 3]
+        )
 
         if deg > 1:
             xx, yy, zz = x * x, y * y, z * z
             xy, yz, xz = x * y, y * z, x * z
-            result = (result +
-                    C2[0] * xy * sh[..., 4] +
-                    C2[1] * yz * sh[..., 5] +
-                    C2[2] * (2.0 * zz - xx - yy) * sh[..., 6] +
-                    C2[3] * xz * sh[..., 7] +
-                    C2[4] * (xx - yy) * sh[..., 8])
+            result = (
+                result
+                + C2[0] * xy * sh[..., 4]
+                + C2[1] * yz * sh[..., 5]
+                + C2[2] * (2.0 * zz - xx - yy) * sh[..., 6]
+                + C2[3] * xz * sh[..., 7]
+                + C2[4] * (xx - yy) * sh[..., 8]
+            )
 
             if deg > 2:
-                result = (result +
-                C3[0] * y * (3 * xx - yy) * sh[..., 9] +
-                C3[1] * xy * z * sh[..., 10] +
-                C3[2] * y * (4 * zz - xx - yy)* sh[..., 11] +
-                C3[3] * z * (2 * zz - 3 * xx - 3 * yy) * sh[..., 12] +
-                C3[4] * x * (4 * zz - xx - yy) * sh[..., 13] +
-                C3[5] * z * (xx - yy) * sh[..., 14] +
-                C3[6] * x * (xx - 3 * yy) * sh[..., 15])
+                result = (
+                    result
+                    + C3[0] * y * (3 * xx - yy) * sh[..., 9]
+                    + C3[1] * xy * z * sh[..., 10]
+                    + C3[2] * y * (4 * zz - xx - yy) * sh[..., 11]
+                    + C3[3] * z * (2 * zz - 3 * xx - 3 * yy) * sh[..., 12]
+                    + C3[4] * x * (4 * zz - xx - yy) * sh[..., 13]
+                    + C3[5] * z * (xx - yy) * sh[..., 14]
+                    + C3[6] * x * (xx - 3 * yy) * sh[..., 15]
+                )
 
                 if deg > 3:
-                    result = (result + C4[0] * xy * (xx - yy) * sh[..., 16] +
-                            C4[1] * yz * (3 * xx - yy) * sh[..., 17] +
-                            C4[2] * xy * (7 * zz - 1) * sh[..., 18] +
-                            C4[3] * yz * (7 * zz - 3) * sh[..., 19] +
-                            C4[4] * (zz * (35 * zz - 30) + 3) * sh[..., 20] +
-                            C4[5] * xz * (7 * zz - 3) * sh[..., 21] +
-                            C4[6] * (xx - yy) * (7 * zz - 1) * sh[..., 22] +
-                            C4[7] * xz * (xx - 3 * yy) * sh[..., 23] +
-                            C4[8] * (xx * (xx - 3 * yy) - yy * (3 * xx - yy)) * sh[..., 24])
+                    result = (
+                        result
+                        + C4[0] * xy * (xx - yy) * sh[..., 16]
+                        + C4[1] * yz * (3 * xx - yy) * sh[..., 17]
+                        + C4[2] * xy * (7 * zz - 1) * sh[..., 18]
+                        + C4[3] * yz * (7 * zz - 3) * sh[..., 19]
+                        + C4[4] * (zz * (35 * zz - 30) + 3) * sh[..., 20]
+                        + C4[5] * xz * (7 * zz - 3) * sh[..., 21]
+                        + C4[6] * (xx - yy) * (7 * zz - 1) * sh[..., 22]
+                        + C4[7] * xz * (xx - 3 * yy) * sh[..., 23]
+                        + C4[8]
+                        * (xx * (xx - 3 * yy) - yy * (3 * xx - yy))
+                        * sh[..., 24]
+                    )
     return result
 
+
 def RGB2SH(rgb):
     return (rgb - 0.5) / C0
 
+
 def SH2RGB(sh):
-    return sh * C0 + 0.5
+    return sh * C0 + 0.5

trellis/representations/gaussian/__init__.py

@@ -1 +1 @@
-from .gaussian_model import Gaussian
+from .gaussian_model import Gaussian

trellis/representations/gaussian/gaussian_model.py

@@ -7,27 +7,27 @@ import utils3d
 
 class Gaussian:
     def __init__(
-            self, 
-            aabb : list,
-            sh_degree : int = 0,
-            mininum_kernel_size : float = 0.0,
-            scaling_bias : float = 0.01,
-            opacity_bias : float = 0.1,
-            scaling_activation : str = "exp",
-            device='cuda'
-        ):
+        self,
+        aabb: list,
+        sh_degree: int = 0,
+        mininum_kernel_size: float = 0.0,
+        scaling_bias: float = 0.01,
+        opacity_bias: float = 0.1,
+        scaling_activation: str = "exp",
+        device="cuda",
+    ):
         self.init_params = {
-            'aabb': aabb,
-            'sh_degree': sh_degree,
-            'mininum_kernel_size': mininum_kernel_size,
-            'scaling_bias': scaling_bias,
-            'opacity_bias': opacity_bias,
-            'scaling_activation': scaling_activation,
+            "aabb": aabb,
+            "sh_degree": sh_degree,
+            "mininum_kernel_size": mininum_kernel_size,
+            "scaling_bias": scaling_bias,
+            "opacity_bias": opacity_bias,
+            "scaling_activation": scaling_activation,
         }
-        
+
         self.sh_degree = sh_degree
         self.active_sh_degree = sh_degree
-        self.mininum_kernel_size = mininum_kernel_size 
+        self.mininum_kernel_size = mininum_kernel_size
         self.scaling_bias = scaling_bias
         self.opacity_bias = opacity_bias
         self.scaling_activation_type = scaling_activation
@@ -48,7 +48,7 @@ class Gaussian:
             actual_covariance = L @ L.transpose(1, 2)
             symm = strip_symmetric(actual_covariance)
             return symm
-        
+
         if self.scaling_activation_type == "exp":
             self.scaling_activation = torch.exp
             self.inverse_scaling_activation = torch.log
@@ -62,74 +62,91 @@ class Gaussian:
         self.inverse_opacity_activation = inverse_sigmoid
 
         self.rotation_activation = torch.nn.functional.normalize
-        
-        self.scale_bias = self.inverse_scaling_activation(torch.tensor(self.scaling_bias)).cuda()
+
+        self.scale_bias = self.inverse_scaling_activation(
+            torch.tensor(self.scaling_bias)
+        ).cuda()
         self.rots_bias = torch.zeros((4)).cuda()
         self.rots_bias[0] = 1
-        self.opacity_bias = self.inverse_opacity_activation(torch.tensor(self.opacity_bias)).cuda()
+        self.opacity_bias = self.inverse_opacity_activation(
+            torch.tensor(self.opacity_bias)
+        ).cuda()
 
     @property
     def get_scaling(self):
         scales = self.scaling_activation(self._scaling + self.scale_bias)
-        scales = torch.square(scales) + self.mininum_kernel_size ** 2
+        scales = torch.square(scales) + self.mininum_kernel_size**2
         scales = torch.sqrt(scales)
         return scales
-    
+
     @property
     def get_rotation(self):
         return self.rotation_activation(self._rotation + self.rots_bias[None, :])
-    
+
     @property
     def get_xyz(self):
         return self._xyz * self.aabb[None, 3:] + self.aabb[None, :3]
-    
+
     @property
     def get_features(self):
-        return torch.cat((self._features_dc, self._features_rest), dim=2) if self._features_rest is not None else self._features_dc
-    
+        return (
+            torch.cat((self._features_dc, self._features_rest), dim=2)
+            if self._features_rest is not None
+            else self._features_dc
+        )
+
     @property
     def get_opacity(self):
         return self.opacity_activation(self._opacity + self.opacity_bias)
-    
-    def get_covariance(self, scaling_modifier = 1):
-        return self.covariance_activation(self.get_scaling, scaling_modifier, self._rotation + self.rots_bias[None, :])
-    
+
+    def get_covariance(self, scaling_modifier=1):
+        return self.covariance_activation(
+            self.get_scaling, scaling_modifier, self._rotation + self.rots_bias[None, :]
+        )
+
     def from_scaling(self, scales):
-        scales = torch.sqrt(torch.square(scales) - self.mininum_kernel_size ** 2)
+        scales = torch.sqrt(torch.square(scales) - self.mininum_kernel_size**2)
         self._scaling = self.inverse_scaling_activation(scales) - self.scale_bias
-        
+
     def from_rotation(self, rots):
         self._rotation = rots - self.rots_bias[None, :]
-    
+
     def from_xyz(self, xyz):
         self._xyz = (xyz - self.aabb[None, :3]) / self.aabb[None, 3:]
-        
+
     def from_features(self, features):
         self._features_dc = features
-        
+
     def from_opacity(self, opacities):
         self._opacity = self.inverse_opacity_activation(opacities) - self.opacity_bias
 
     def construct_list_of_attributes(self):
-        l = ['x', 'y', 'z', 'nx', 'ny', 'nz']
+        l = ["x", "y", "z", "nx", "ny", "nz"]
         # All channels except the 3 DC
-        for i in range(self._features_dc.shape[1]*self._features_dc.shape[2]):
-            l.append('f_dc_{}'.format(i))
-        l.append('opacity')
+        for i in range(self._features_dc.shape[1] * self._features_dc.shape[2]):
+            l.append("f_dc_{}".format(i))
+        l.append("opacity")
         for i in range(self._scaling.shape[1]):
-            l.append('scale_{}'.format(i))
+            l.append("scale_{}".format(i))
         for i in range(self._rotation.shape[1]):
-            l.append('rot_{}'.format(i))
+            l.append("rot_{}".format(i))
         return l
-        
+
     def save_ply(self, path, transform=[[1, 0, 0], [0, 0, -1], [0, 1, 0]]):
         xyz = self.get_xyz.detach().cpu().numpy()
         normals = np.zeros_like(xyz)
-        f_dc = self._features_dc.detach().transpose(1, 2).flatten(start_dim=1).contiguous().cpu().numpy()
+        f_dc = (
+            self._features_dc.detach()
+            .transpose(1, 2)
+            .flatten(start_dim=1)
+            .contiguous()
+            .cpu()
+            .numpy()
+        )
         opacities = inverse_sigmoid(self.get_opacity).detach().cpu().numpy()
         scale = torch.log(self.get_scaling).detach().cpu().numpy()
         rotation = (self._rotation + self.rots_bias[None, :]).detach().cpu().numpy()
-        
+
         if transform is not None:
             transform = np.array(transform)
             xyz = np.matmul(xyz, transform.T)
@@ -137,20 +154,29 @@ class Gaussian:
             rotation = np.matmul(transform, rotation)
             rotation = utils3d.numpy.matrix_to_quaternion(rotation)
 
-        dtype_full = [(attribute, 'f4') for attribute in self.construct_list_of_attributes()]
+        dtype_full = [
+            (attribute, "f4") for attribute in self.construct_list_of_attributes()
+        ]
 
         elements = np.empty(xyz.shape[0], dtype=dtype_full)
-        attributes = np.concatenate((xyz, normals, f_dc, opacities, scale, rotation), axis=1)
+        attributes = np.concatenate(
+            (xyz, normals, f_dc, opacities, scale, rotation), axis=1
+        )
         elements[:] = list(map(tuple, attributes))
-        el = PlyElement.describe(elements, 'vertex')
+        el = PlyElement.describe(elements, "vertex")
         PlyData([el]).write(path)
 
     def load_ply(self, path, transform=[[1, 0, 0], [0, 0, -1], [0, 1, 0]]):
         plydata = PlyData.read(path)
 
-        xyz = np.stack((np.asarray(plydata.elements[0]["x"]),
-                        np.asarray(plydata.elements[0]["y"]),
-                        np.asarray(plydata.elements[0]["z"])),  axis=1)
+        xyz = np.stack(
+            (
+                np.asarray(plydata.elements[0]["x"]),
+                np.asarray(plydata.elements[0]["y"]),
+                np.asarray(plydata.elements[0]["z"]),
+            ),
+            axis=1,
+        )
         opacities = np.asarray(plydata.elements[0]["opacity"])[..., np.newaxis]
 
         features_dc = np.zeros((xyz.shape[0], 3, 1))
@@ -159,43 +185,65 @@ class Gaussian:
         features_dc[:, 2, 0] = np.asarray(plydata.elements[0]["f_dc_2"])
 
         if self.sh_degree > 0:
-            extra_f_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("f_rest_")]
-            extra_f_names = sorted(extra_f_names, key = lambda x: int(x.split('_')[-1]))
-            assert len(extra_f_names)==3*(self.sh_degree + 1) ** 2 - 3
+            extra_f_names = [
+                p.name
+                for p in plydata.elements[0].properties
+                if p.name.startswith("f_rest_")
+            ]
+            extra_f_names = sorted(extra_f_names, key=lambda x: int(x.split("_")[-1]))
+            assert len(extra_f_names) == 3 * (self.sh_degree + 1) ** 2 - 3
             features_extra = np.zeros((xyz.shape[0], len(extra_f_names)))
             for idx, attr_name in enumerate(extra_f_names):
                 features_extra[:, idx] = np.asarray(plydata.elements[0][attr_name])
             # Reshape (P,F*SH_coeffs) to (P, F, SH_coeffs except DC)
-            features_extra = features_extra.reshape((features_extra.shape[0], 3, (self.max_sh_degree + 1) ** 2 - 1))
+            features_extra = features_extra.reshape(
+                (features_extra.shape[0], 3, (self.max_sh_degree + 1) ** 2 - 1)
+            )
 
-        scale_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("scale_")]
-        scale_names = sorted(scale_names, key = lambda x: int(x.split('_')[-1]))
+        scale_names = [
+            p.name
+            for p in plydata.elements[0].properties
+            if p.name.startswith("scale_")
+        ]
+        scale_names = sorted(scale_names, key=lambda x: int(x.split("_")[-1]))
         scales = np.zeros((xyz.shape[0], len(scale_names)))
         for idx, attr_name in enumerate(scale_names):
             scales[:, idx] = np.asarray(plydata.elements[0][attr_name])
 
-        rot_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("rot")]
-        rot_names = sorted(rot_names, key = lambda x: int(x.split('_')[-1]))
+        rot_names = [
+            p.name for p in plydata.elements[0].properties if p.name.startswith("rot")
+        ]
+        rot_names = sorted(rot_names, key=lambda x: int(x.split("_")[-1]))
         rots = np.zeros((xyz.shape[0], len(rot_names)))
         for idx, attr_name in enumerate(rot_names):
             rots[:, idx] = np.asarray(plydata.elements[0][attr_name])
-            
+
         if transform is not None:
             transform = np.array(transform)
             xyz = np.matmul(xyz, transform)
             rotation = utils3d.numpy.quaternion_to_matrix(rotation)
             rotation = np.matmul(rotation, transform)
             rotation = utils3d.numpy.matrix_to_quaternion(rotation)
-            
+
         # convert to actual gaussian attributes
         xyz = torch.tensor(xyz, dtype=torch.float, device=self.device)
-        features_dc = torch.tensor(features_dc, dtype=torch.float, device=self.device).transpose(1, 2).contiguous()
+        features_dc = (
+            torch.tensor(features_dc, dtype=torch.float, device=self.device)
+            .transpose(1, 2)
+            .contiguous()
+        )
         if self.sh_degree > 0:
-            features_extra = torch.tensor(features_extra, dtype=torch.float, device=self.device).transpose(1, 2).contiguous()
-        opacities = torch.sigmoid(torch.tensor(opacities, dtype=torch.float, device=self.device))
+            features_extra = (
+                torch.tensor(features_extra, dtype=torch.float, device=self.device)
+                .transpose(1, 2)
+                .contiguous()
+            )
+        opacities = torch.sigmoid(
+            torch.tensor(opacities, dtype=torch.float, device=self.device)
+        )
         scales = torch.exp(torch.tensor(scales, dtype=torch.float, device=self.device))
         rots = torch.tensor(rots, dtype=torch.float, device=self.device)
-        
+
         # convert to _hidden attributes
         self._xyz = (xyz - self.aabb[None, :3]) / self.aabb[None, 3:]
         self._features_dc = features_dc
@@ -204,6 +252,10 @@ class Gaussian:
         else:
             self._features_rest = None
         self._opacity = self.inverse_opacity_activation(opacities) - self.opacity_bias
-        self._scaling = self.inverse_scaling_activation(torch.sqrt(torch.square(scales) - self.mininum_kernel_size ** 2)) - self.scale_bias
+        self._scaling = (
+            self.inverse_scaling_activation(
+                torch.sqrt(torch.square(scales) - self.mininum_kernel_size**2)
+            )
+            - self.scale_bias
+        )
         self._rotation = rots - self.rots_bias[None, :]
-        

trellis/representations/gaussian/general_utils.py

@@ -3,7 +3,7 @@
 # GRAPHDECO research group, https://team.inria.fr/graphdeco
 # All rights reserved.
 #
-# This software is free for non-commercial, research and evaluation use 
+# This software is free for non-commercial, research and evaluation use
 # under the terms of the LICENSE.md file.
 #
 # For inquiries contact  [email protected]
@@ -15,8 +15,10 @@ from datetime import datetime
 import numpy as np
 import random
 
+
 def inverse_sigmoid(x):
-    return torch.log(x/(1-x))
+    return torch.log(x / (1 - x))
+
 
 def PILtoTorch(pil_image, resolution):
     resized_image_PIL = pil_image.resize(resolution)
@@ -26,6 +28,7 @@ def PILtoTorch(pil_image, resolution):
     else:
         return resized_image.unsqueeze(dim=-1).permute(2, 0, 1)
 
+
 def get_expon_lr_func(
     lr_init, lr_final, lr_delay_steps=0, lr_delay_mult=1.0, max_steps=1000000
 ):
@@ -61,6 +64,7 @@ def get_expon_lr_func(
 
     return helper
 
+
 def strip_lowerdiag(L):
     uncertainty = torch.zeros((L.shape[0], 6), dtype=torch.float, device="cuda")
 
@@ -72,45 +76,52 @@ def strip_lowerdiag(L):
     uncertainty[:, 5] = L[:, 2, 2]
     return uncertainty
 
+
 def strip_symmetric(sym):
     return strip_lowerdiag(sym)
 
+
 def build_rotation(r):
-    norm = torch.sqrt(r[:,0]*r[:,0] + r[:,1]*r[:,1] + r[:,2]*r[:,2] + r[:,3]*r[:,3])
+    norm = torch.sqrt(
+        r[:, 0] * r[:, 0] + r[:, 1] * r[:, 1] + r[:, 2] * r[:, 2] + r[:, 3] * r[:, 3]
+    )
 
     q = r / norm[:, None]
 
-    R = torch.zeros((q.size(0), 3, 3), device='cuda')
+    R = torch.zeros((q.size(0), 3, 3), device="cuda")
 
     r = q[:, 0]
     x = q[:, 1]
     y = q[:, 2]
     z = q[:, 3]
 
-    R[:, 0, 0] = 1 - 2 * (y*y + z*z)
-    R[:, 0, 1] = 2 * (x*y - r*z)
-    R[:, 0, 2] = 2 * (x*z + r*y)
-    R[:, 1, 0] = 2 * (x*y + r*z)
-    R[:, 1, 1] = 1 - 2 * (x*x + z*z)
-    R[:, 1, 2] = 2 * (y*z - r*x)
-    R[:, 2, 0] = 2 * (x*z - r*y)
-    R[:, 2, 1] = 2 * (y*z + r*x)
-    R[:, 2, 2] = 1 - 2 * (x*x + y*y)
+    R[:, 0, 0] = 1 - 2 * (y * y + z * z)
+    R[:, 0, 1] = 2 * (x * y - r * z)
+    R[:, 0, 2] = 2 * (x * z + r * y)
+    R[:, 1, 0] = 2 * (x * y + r * z)
+    R[:, 1, 1] = 1 - 2 * (x * x + z * z)
+    R[:, 1, 2] = 2 * (y * z - r * x)
+    R[:, 2, 0] = 2 * (x * z - r * y)
+    R[:, 2, 1] = 2 * (y * z + r * x)
+    R[:, 2, 2] = 1 - 2 * (x * x + y * y)
     return R
 
+
 def build_scaling_rotation(s, r):
     L = torch.zeros((s.shape[0], 3, 3), dtype=torch.float, device="cuda")
     R = build_rotation(r)
 
-    L[:,0,0] = s[:,0]
-    L[:,1,1] = s[:,1]
-    L[:,2,2] = s[:,2]
+    L[:, 0, 0] = s[:, 0]
+    L[:, 1, 1] = s[:, 1]
+    L[:, 2, 2] = s[:, 2]
 
     L = R @ L
     return L
 
+
 def safe_state(silent):
     old_f = sys.stdout
+
     class F:
         def __init__(self, silent):
             self.silent = silent
@@ -118,7 +129,14 @@ def safe_state(silent):
         def write(self, x):
             if not self.silent:
                 if x.endswith("\n"):
-                    old_f.write(x.replace("\n", " [{}]\n".format(str(datetime.now().strftime("%d/%m %H:%M:%S")))))
+                    old_f.write(
+                        x.replace(
+                            "\n",
+                            " [{}]\n".format(
+                                str(datetime.now().strftime("%d/%m %H:%M:%S"))
+                            ),
+                        )
+                    )
                 else:
                     old_f.write(x)