For the first time, we introduce Content-Preserving Style Transfer functionality to Qwen-Image-Edit, which supports transferring various style cues from style reference to content reference while preserving the characteristics of content reference in high efficiency, i.e. 4 sampling steps.

Please note that our style transfer model is based on Qwen-Image-Edit-2509, and has to be used with Qwen-Image-Lightning Lora, which we have converted to Diffsynth format for compatibility. Otherwise, the model may suffer from either low-speed or low-quality. Our github page is QwenStyle.

Install DiffSynth

git clone https://github.com/modelscope/DiffSynth-Studio.git  
cd DiffSynth-Studio
pip install -e .

Please download our style transfer lora and lightning lora from this link

Then run infer_style_transfer.py for inference. We have tested the model on one H100, which takes 5 seconds to generate the result.

Our training framework is based on DiffSynth-Studio. Special thanks to the authors of DiffSynth.

We will open-source all our training data if the stars exceed 200.

We release the tech report of QwenStyle V1. We are keep refining QwenStyle and will update new versions in the future. Please light a star for our project and cite our work if you find it helpful.