# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

This is a FastAPI-based virtual try-on service that integrates with ComfyUI for AI-powered clothing try-on. The service accepts base64-encoded images (model photo, shirt, pants) and returns a generated try-on result.

## Running the Service

Start the server:
```bash
python service2.py
```

The server runs on `http://0.0.0.0:47698` by default.

## Dependencies

Required Python packages (install via pip):
- fastapi
- uvicorn
- httpx
- pydantic

## Architecture

### Request Flow
1. Client sends POST to `/try-on` with base64 images (model_image, shirt_image, pants_image)
2. Server decodes images and uploads to ComfyUI at `http://127.0.0.1:8188`
3. Server modifies workflow JSON (change2_1_ex.json) with uploaded image filenames
4. Workflow submitted to ComfyUI queue
5. Server polls ComfyUI history endpoint until completion (max 300s timeout)
6. Result image downloaded from ComfyUI and returned as base64

### Key Components

**service2.py**: Main FastAPI application
- `/try-on`: Main endpoint for virtual try-on
- `/health`: Health check endpoint
- `/`: Serves static frontend (index.html)

**change2_1_ex.json**: ComfyUI workflow configuration
- Node "11": Model image input (LoadImage)
- Node "10": Shirt image input (LoadImage)
- Node "33": Output image (SaveImage)
- Node "31": Text prompt for image editing

**static/index.html**: Frontend interface for testing

### ComfyUI Integration

The service expects ComfyUI to be running at `http://127.0.0.1:8188` with:
- FireRed-Image-Edit model loaded
- Qwen 2.5 VL CLIP model
- Required custom nodes for QwenEditConfigPreparer, QwenEditOutputExtractor

### Workflow Modification

When processing requests, the code dynamically updates these workflow nodes:
- `workflow["11"]["inputs"]["image"]`: Model photo filename
- `workflow["10"]["inputs"]["image"]`: Shirt photo filename
- `workflow["31"]["inputs"]["value"]`: Text prompt (currently hardcoded in Chinese)

## Important Notes

- Images are uploaded with unique filenames (UUID prefix) to avoid caching issues
- The pants_image parameter is currently unused (commented out in code)
- Workflow uses 8 sampling steps with euler sampler and beta scheduler
- Output resolution is 768x1024 pixels
- CORS is enabled for all origins