---
license: apache-2.0
tags:
- video-generation
- video-editing
- in-context-learning
- pytorch
pipeline_tag: video-to-video
library_name: transformers
authors:
- XiangpengYang
- horizonwind2004
---

<div align="center">

  <h1 style="margin: 0; font-size: 1.8em;">
    Unified Video Editing with Temporal Reasoner
  </h1>

  <h4 style="margin: 15px 0; color: #2c3e50;">
    👁️ See &rarr; 🧠 Reason &rarr; ✏️ Edit
  </h4>

  <h4 style="margin: 15px 0; color: #2c3e50;">
    🚀 A Chain of Frames editing method enbale temporal reasoning and 4x video length generalization with just 50k training pairs!
  </h4>

  <a href="https://huggingface.co/papers/2512.07469"><img src="https://img.shields.io/badge/HuggingFace-Daily_Paper-ffd21e.svg" alt="Daily Paper"></a>
  <a href="https://arxiv.org/abs/2512.07469"><img src="https://img.shields.io/badge/arXiv-2512.07469-b31b1b.svg" alt="arXiv"></a>
  <a href="https://videocof.github.io"><img src="https://img.shields.io/badge/Project-Page-green" alt="Project Page"></a>
  <a href="https://github.com/knightyxp/VideoCoF"><img src="https://img.shields.io/badge/GitHub-Repo-blue?logo=github" alt="GitHub"></a>

</div>

<div align="center">
  <b>
    <a href="https://scholar.google.com/citations?user=reiIeYMAAAAJ">Xiangpeng Yang</a><sup>1</sup>,
    <a href="https://horizonwind2004.github.io/">Ji Xie</a><sup>2</sup>,
    <a href="https://scholar.google.com/citations?user=OvfI_HMAAAAJ">Yiyuan Yang</a><sup>1</sup>,
    <a href="https://scholar.google.com/citations?user=zfeWd6gAAAAJ">Yan Huang</a><sup>1</sup>,
    <a href="https://scholar.google.com/citations?user=sCuACdkAAAAJ">Min Xu</a><sup>1</sup>,
    <a href="https://scholar.google.com/citations?user=sCuACdkAAAAJ">Qiang Wu</a><sup>1</sup>
  </b>
  <br>
  <span style="font-size: 1em; color: #555;"><sup>1</sup>University of Technology Sydney, <sup>2</sup>Zhejiang University</span>
</div>

<br>

# VideoCoF: Unified Video Editing with Temporal Reasoner


**VideoCoF** is a unified video editing model that bridges the gap between expert models (precise but restricted) and unified in-context models (flexible but spatially inaccurate). By introducing a **"See &rarr; Reason &rarr; Edit"**, a Chain-of-Frames paradigm, VideoCoF predicts reasoning tokens before generating the target video tokens, thereby removing the need for user-provided masks while achieving precise instruction to-region alignment. 

<div align="center">
  <a href="https://www.youtube.com/watch?v=XrYj0Qmc49w" target="_blank">
    <img src="https://img.youtube.com/vi/XrYj0Qmc49w/maxresdefault.jpg" 
         alt="Video Demo" 
         width="80%" 
         style="max-width:900px; border-radius:10px; box-shadow:0 0 10px rgba(0,0,0,0.15);">
  </a>
  <br>
  <em>Click the image above to watch the full video on YouTube 🎬</em>
</div>

## 🌟 Key Capabilities
![](assets/motivation_v2.gif)

1.  **Temporal Reasoning**: Adopts a unique approach where the model first identifies *where* and *how* to edit (Reasoning) before predicting the target video tokens.
2.  **Data Efficiency**: Achieves SOTA performance with only **50k training pairs** (33 frames each).
3.  **Length Extrapolation**: Demonstrates robust multi-shot editing and can generalize to videos **4&times; longer** than training samples.
4.  **Versatile Editing**: Supports:
    * Object Removal
    * Object Addition
    * Object Swap
    * Local Style Transfer

## 🔧 Quick Start

To use these weights, please refer to the official [GitHub Repository](https://github.com/knightyxp/VideoCoF) for inference code and environment setup.

### Installation

```bash
git clone https://github.com/knightyxp/VideoCoF
cd VideoCoF

# 1. Create and activate a conda environment
conda create -n videocof python=3.10
conda activate videocof

# 2. Install PyTorch (Choose version compatible with your CUDA)
# For standard GPUs (CUDA 12.1):
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121

# For Hopper GPUs (e.g., H100/H800) requiring fast inference:
# pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128

# 3. Install other dependencies
pip install -r requirements.txt
```

**Note on Flash Attention:**
We recommend using **FlashAttention-3** (currently beta) for optimal performance, especially on NVIDIA H100/H800 GPUs. 
If you are using these GPUs, please follow the [official FlashAttention-3 installation guide](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#flashattention-3-beta-release) after installing the compatible PyTorch version (e.g., PyTorch 2.8 + CUDA 12.8).

### Download Models

*   **Wan-2.1-T2V-14B Pretrained Weights:**
    
    ```bash
    git lfs install
    git clone https://huggingface.co/Wan-AI/Wan2.1-T2V-14B
    
    # Or using huggingface-cli:
    # hf download Wan-AI/Wan2.1-T2V-14B --local-dir Wan2.1-T2V-14B
    ```

*   **VideoCoF Checkpoint:**
    
    ```bash
    git lfs install
    git clone https://huggingface.co/XiangpengYang/VideoCoF videocof_weight

    # Or using huggingface-cli:
    # hf download XiangpengYang/VideoCoF --local-dir videocof_weight
    ```

### Inference

```bash
export CUDA_VISIBLE_DEVICES=0
torchrun --nproc_per_node=1 inference.py \
  --video_path assets/two_man.mp4 \
  --prompt "Remove the young man with short black hair wearing black shirt on the left." \
  --output_dir results/obj_rem \
  --model_name /scratch3/yan204/models/Wan2.1-T2V-14B \
  --seed 0 \
  --num_frames 33 \
  --source_frames 33 \
  --reasoning_frames 4 \
  --repeat_rope \
  --videocof_path videocof_weight/videocof.safetensors
```

For parallel inference:

```bash
sh scripts/parallel_infer.sh
```

## 🙏 Acknowledgments

We thank the authors of related works and the open-source community [VideoX-Fun](https://github.com/aigc-apps/VideoX-Fun) and [Wan](https://github.com/Wan-Video/Wan2.1) for their contributions.

## 📜 License

This project is licensed under the [Apache License 2.0](LICENSE).

## 📮 Contact

For any questions, please feel free to reach out to the author Xiangpeng Yang [@knightyxp](https://github.com/knightyxp), email: knightyxp@gmail.com/Xiangpeng.Yang@student.uts.edu.au

## 📄 Citation

If you find this work useful for your research, please consider citing:

```bibtex
@article{yang2025videocof,
  title={Unified Video Editing with Temporal Reasoner},
  author={Yang, Xiangpeng and Xie, Ji and Yang, Yiyuan and Huang, Yan and Xu, Min and Wu, Qiang},
  journal={arXiv preprint arXiv:2512.07469},
  year={2025}
}
```

<div align="center">
  ❤️ **If you find this project helpful, please consider giving it a like!** ❤️
</div>