You can deploy this model and start earning money today!

ReyChiaro/
StyleController

diffusers

style-transfer

image-editing

controllable-generation

ReyChiaro/SmoothStyle

2608.08125

Qwen/Qwen-Image-Edit-2509

adapter

apache-2.0

Views

No views yet

StyleController

Staying True to the Origin: Continuous Image Stylization with Smooth Transitions

Rui Xu · Hanmo Zhang · Songhua Liu
Shanghai Jiao Tong University

Project Page

arXiv

GitHub

StyleController is a controllable image-stylization adapter collection. Given a content image, a style reference, a text instruction, and a scalar strength, it generates stylized images while providing continuous control over the amount of transferred style.

This repository contains model weights only. Inference and training code are maintained in the StyleController GitHub repository.

Available models

Variant	Base model	Precision	Strength anchors	Status
`qwenimage_edit_2509`	`Qwen/Qwen-Image-Edit-2509`	BF16	0.1–0.9	Available

Repository layout

text

StyleController/
├── README.md
├── MODEL_CARD.md
├── LICENSE
├── NOTICE
├── model_manifest.json
└── qwenimage_edit_2509/
    ├── README.md
    ├── stylecontroller_config.json
    ├── checksums.sha256
    ├── styctrl_qwenimage_edit_2509.safetensors
    └── projectors/
        ├── s01.safetensors
        ├── ...
        └── s09.safetensors

Download

Install the Hugging Face CLI and download the current variant:

bash

pip install -U huggingface_hub

hf download ReyChiaro/StyleController \
  --include "qwenimage_edit_2509/*" \
  --local-dir hf_weights

Inference

Clone and install the implementation:

bash

git clone https://github.com/ReyChiaro/StyleController.git
cd StyleController
uv sync

Run Qwen-Image-Edit-2509 inference at strength 0.55:

bash

uv run python infer_styctrl.py \
  --model qwen \
  --pretrained-pipeline Qwen/Qwen-Image-Edit-2509 \
  --content examples/content.jpg \
  --style examples/style.jpg \
  --prompt "Transfer the style of image 2 to image 1 while preserving the content structure." \
  --strength 0.55 \
  --output outputs/result_055.png \
  --lora-checkpoint hf_weights/qwenimage_edit_2509/styctrl_qwenimage_edit_2509.safetensors \
  --projector-checkpoints \
    hf_weights/qwenimage_edit_2509/projectors/s01.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s02.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s03.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s04.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s05.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s06.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s07.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s08.safetensors \
    hf_weights/qwenimage_edit_2509/projectors/s09.safetensors \
  --anchor-strengths 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

License

The model-weight release is provided under the Apache License 2.0. The StyleController source code is separately licensed under MIT. The Qwen base model and SmoothStyle dataset remain governed by their own licenses and notices; see NOTICE.

Citation

bibtex

@misc{xu2026stayingtrueorigincontinuous,
      title={Staying True to the Origin: Continuous Image Stylization with Smooth Transitions},
      author={Rui Xu and Hanmo Zhang and Songhua Liu},
      year={2026},
      eprint={2608.08125},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.08125},
}