Summary

Edit-R1 enhances image editing by fine-tuning diffusion models with "Diffusion Negative-Aware Finetuning" and implicit feedback from Multimodal Large Language Models (MLLMs). This approach aims for precise control and higher quality AI-driven image manipulation, targeting researchers and developers.

How It Works

Leveraging the DiffusionNFT codebase, Edit-R1 fine-tunes diffusion models using a training-free reward model derived from pretrained MLLMs. This negative-aware feedback mechanism guides the fine-tuning process, improving the model's understanding of undesirable edits for more robust and accurate results.

Quick Start & Requirements

Installation requires cloning the repo, creating a Python 3.10.16 Conda environment, and running pip install -e .. Training necessitates deploying a reward server (python reward_server/reward_server.py) and configuring the REWARD_SERVER environment variable. Data requires a specific directory structure with images and metadata JSONL files.

• Python: 3.10.16
• Dependencies: Conda, pip
• Links: Hugging Face Collection, UniWorld-V2 Arxiv, UniWorld-V1 Arxiv, ImgEdit Arxiv.

Highlighted Details

• Open-sourced models include UniWorld-Qwen-Image-Edit-2509 and UniWorld-FLUX.1-Kontext-Dev.
• Demonstrated capabilities cover precise object manipulation, style transfer, scene composition, and detailed texture overlays, as evidenced by case comparisons.

Maintenance & Community

The project references Arxiv papers and a Hugging Face collection. No direct links to community forums or a public roadmap are provided in the README.

Licensing & Compatibility

The primary license is Apache. However, FLUX weights are under a "FLUX.1 [dev] Non-Commercial License," restricting their use in commercial applications.

Limitations & Caveats

The primary limitation is the non-commercial use restriction for FLUX weights. The README does not detail other potential limitations, unsupported platforms, or known bugs.