Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

QwenStyle: Content-Preserving Style Transfer with Qwen-Image-Edit

Shiwen Zhang, Haibin Huang, Chi Zhang, Xuelong Li

TL;DR

QwenStyle addresses the challenge of content-preserving style transfer within Diffusion Transformer-based frameworks by leveraging Qwen-Image-Edit. It introduces a three-stage Curriculum Continual Learning approach built on two triplet datasets, D_collected and D_synthetic, to simultaneously preserve content and generalize to unseen styles. The model uses Lora adapters and a fixed prompting scheme, achieving state-of-the-art performance on style similarity, content preservation, and aesthetics. The work demonstrates strong practical potential for controllable, content-faithful style editing in open-source image editing pipelines.

Abstract

Content-Preserving Style transfer, given content and style references, remains challenging for Diffusion Transformers (DiTs) due to its internal entangled content and style features. In this technical report, we propose the first content-preserving style transfer model trained on Qwen-Image-Edit, which activates Qwen-Image-Edit's strong content preservation and style customization capability. We collected and filtered high quality data of limited specific styles and synthesized triplets with thousands categories of style images in-the-wild. We introduce the Curriculum Continual Learning framework to train QwenStyle with such mixture of clean and noisy triplets, which enables QwenStyle to generalize to unseen styles without degradation of the precise content preservation capability. Our QwenStyle V1 achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.

QwenStyle: Content-Preserving Style Transfer with Qwen-Image-Edit

TL;DR

QwenStyle addresses the challenge of content-preserving style transfer within Diffusion Transformer-based frameworks by leveraging Qwen-Image-Edit. It introduces a three-stage Curriculum Continual Learning approach built on two triplet datasets, D_collected and D_synthetic, to simultaneously preserve content and generalize to unseen styles. The model uses Lora adapters and a fixed prompting scheme, achieving state-of-the-art performance on style similarity, content preservation, and aesthetics. The work demonstrates strong practical potential for controllable, content-faithful style editing in open-source image editing pipelines.

Abstract

Content-Preserving Style transfer, given content and style references, remains challenging for Diffusion Transformers (DiTs) due to its internal entangled content and style features. In this technical report, we propose the first content-preserving style transfer model trained on Qwen-Image-Edit, which activates Qwen-Image-Edit's strong content preservation and style customization capability. We collected and filtered high quality data of limited specific styles and synthesized triplets with thousands categories of style images in-the-wild. We introduce the Curriculum Continual Learning framework to train QwenStyle with such mixture of clean and noisy triplets, which enables QwenStyle to generalize to unseen styles without degradation of the precise content preservation capability. Our QwenStyle V1 achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.
Paper Structure (14 sections, 1 equation, 3 figures, 1 table, 1 algorithm)

This paper contains 14 sections, 1 equation, 3 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. QwenStyle
  3. Triplet Training Dataset Construction
  4. Content-Preserving Style Transfer with Curriculum Continual Learning
  5. Training and Inference
  6. Experiments
  7. Implementation Details.
  8. Evaluation Benchmark.
  9. Evaluation Metrics.
  10. Comparison with State-of-the-art Methods
  11. Quantitative Comparison
  12. Qualitative Comparison
  13. Limitations.
  14. Conclusion

Figures (3)

  • Figure 1: QwenStyle accepts style and content references for content-preserving style transfer, while maintaining high aesthetics merit. QwenStyle is the first content-preserving style transfer model built on Qwen-Image.
  • Figure 2: Collected triplets $D_{collected}$ and synthetic triplets $D_{synthetic}$
  • Figure 3: Qualitative Comparison with State-of-the-art Style Transfer Models.