app.py

import gradio as gr
from diffusers import DiffusionPipeline
import torch
from flask import Flask, request, jsonify
import threading
import io
import base64
from PIL import Image

# Carregamento do pipeline do modelo
pipe = DiffusionPipeline.from_pretrained(
    "HiDream-ai/HiDream-E1-Full",
    torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")

# Flask app para API
app = Flask(__name__)

def image_to_base64(image: Image.Image) -> str:
    buffered = io.BytesIO()
    image.save(buffered, format="PNG")
    return base64.b64encode(buffered.getvalue()).decode()

@app.route("/v1/texttoimage/completions", methods=["POST"])
def text_to_image():
    data = request.json
    prompt = data.get("prompt", "").strip()
    if not prompt:
        return jsonify({"error": "Prompt vazio"}), 400
    result = pipe(prompt)
    img_b64 = image_to_base64(result.images[0])
    return jsonify({"image_base64": img_b64})

# Executa Flask em uma thread separada
def run_flask():
    app.run(host="0.0.0.0", port=7860)

threading.Thread(target=run_flask, daemon=True).start()

# Gradio interface
def gerar_imagem_gradio(prompt):
    result = pipe(prompt)
    return result.images[0]

with gr.Blocks() as demo:
    gr.Markdown("## Chat Text-to-Image com API Flask integrada")
    gr.Markdown("API disponível em: `http://localhost:7860/v1/texttoimage/completions`")
    
    chat = gr.Chatbot()
    txt = gr.Textbox(placeholder="Digite seu prompt aqui e pressione Enter", show_label=False)

    def responder(prompt, chat_history):
        img = gerar_imagem_gradio(prompt)
        chat_history = chat_history + [(prompt, img)]
        return chat_history, ""

    txt.submit(responder, inputs=[txt, chat], outputs=[chat, txt])

demo.launch(server_name="0.0.0.0", server_port=7860)