LoRA adapter for Qwen-Image-Edit-2511, trained with Flow-GRPO and a product consistency reward. Plug this into Qwen/Qwen-Image-Edit-2511 to get a model that better preserves product identity (color, shape, logo) during image editing.

Base model: Qwen/Qwen-Image-Edit-2511
Training: GRPO (Group Relative Policy Optimization) on e-commerce product editing data
Reward: Qwen3-VL-30B-A3B-Instruct as VLM judge, 1-5 product consistency score
Result: Reward improved from 0.022 → 0.480 (21.8× over SFT baseline)
LoRA rank / alpha: 64 / 128
Checkpoint: Epoch 24 (peak performance)

Usage

python

from peft import PeftModel
from huggingface_hub import snapshot_download

# Load base model (via DiffSynth-Studio)
# See https://github.com/Matteoooo46/ReCoEdit for full inference script
lora_path = snapshot_download("Matteoooo46/ReCoEdit-RL")
pipe.load_lora(pipe.dit, lora_path + "/adapter_model.safetensors")

See the ReCoEdit repository for the complete inference pipeline including APG guidance and prompt rewriter.

Citation

bibtex

@misc{recoedit2025,
  title={ReCoEdit: Rewriter-Guided Consistency Alignment for Product Image Editing},
  author={},
  year={2025},
  howpublished={\url{https://github.com/Matteoooo46/ReCoEdit}},
}