Matteoooo46/ReCoEdit-rewriter

ReCoEdit — Prompt Rewriter LoRA

LoRA adapter for Qwen3-VL-8B-Instruct, trained to rewrite raw e-commerce product prompts into structured editing instructions for Qwen-Image-Edit.

• Base model: Qwen/Qwen3-VL-8B-Instruct
• Training: SFT via LLaMA-Factory on paired (product image, raw prompt → structured prompt) data
• Input: Product reference image + raw editing prompt (Chinese)
• Output: Structured editing prompt optimized for Qwen-Image-Edit (120-150 chars, Chinese)
• LoRA rank: 64, trained for 10,000 steps

Usage

from transformers import AutoProcessor
from qwen_vl_utils import process_vision_info
from peft import PeftModel
from transformers import Qwen3VLForConditionalGeneration
from huggingface_hub import snapshot_download

base_model = "Qwen/Qwen3-VL-8B-Instruct"
lora_path = snapshot_download("Matteoooo46/ReCoEdit-rewriter")

model = Qwen3VLForConditionalGeneration.from_pretrained(
    base_model, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(model, lora_path)
processor = AutoProcessor.from_pretrained(base_model, trust_remote_code=True)

See the ReCoEdit repository for the complete inference pipeline.

Citation

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