README.md

metadata

title: StyleController
emoji: 🎨
colorFrom: purple
colorTo: indigo
sdk: gradio
sdk_version: 6.15.1
app_file: app.py
short_description: Continuous image stylization with strength control
python_version: '3.12'
startup_duration_timeout: 1h

StyleController

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

This Space demonstrates StyleController, a controllable image stylization framework built on Qwen-Image-Edit-2509. Given a content image and a style reference, it generates high-fidelity stylized images with explicit control over style strength via a slider.

How it works

1. Provide a content image (the image you want stylized).
2. Provide a style reference image (the visual style to transfer).
3. Adjust the style strength slider (0 = no style, 1 = full stylization).
4. Click Stylize to generate the output.

The model combines a full-strength LoRA adapter with lightweight projector anchors learned at discrete style strengths. B-spline interpolation between these anchors produces smooth, continuous transitions while preserving content structure.

Technical details

• Base model: Qwen/Qwen-Image-Edit-2509
• Adapter: ReyChiaro/StyleController
• LoRA rank: 128, layers 0–59
• Projector type: low-rank linear
• Interpolation: B-spline (order 3, clamped knot vector)
• Inference: 16 steps, bf16, seed 42

Citation

@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},
}

License

StyleController source code is MIT-licensed. Adapter weights are Apache-2.0.

Example image credits

Content images are from linoyts/repo-to-space-example-inputs (CC0-1.0).

Style-reference images each represent a distinct visual style category and are drawn from public-domain / Creative-Commons sources:

• Classical oil painting — Claude Monet, Wild Poppies Near Argenteuil (1873), Musée d'Orsay, via Google Arts & Culture / Wikimedia Commons. Public domain.
• Watercolor — John Singer Sargent, Muddy Alligators (1917), Worcester Art Museum, via Wikimedia Commons. Public domain.
• 3D render / CGI — Classroom Blender demo scene by Christophe Seux (blender.org), via Wikimedia Commons. CC0-1.0.
• Flat design / vector illustration — Vector Illustration of a Mountain Landscape with Pine Trees by Missaruthra, via Wikimedia Commons. CC0-1.0.
• Anime / cel illustration — flat anime-style illustration from linoyts/repo-to-space-example-inputs (CC0-1.0).