Publications
カンファレンス (国内) SCAdapter: A Content-Style Disentanglement Approach for Diffusion-Based Style Transfer
Luan Thanh Trinh, 土井 賢治, 長内 淳樹
第28回 画像の認識・理解シンポジウム MIRU2025 (MIRU2025)
2025.7.30
Style transfer in diffusion models is primarily achieved through self-attention manipulation or textual inversion, each with limitations. Self-attention excels in various scenarios but struggles with photo-to-photo transfers due to its inability to effectively represent photorealistic styles. Textual inversion performs better in photo-to-photo tasks but results in weaker style influences and fails when original and target styles differ significantly. Both methods leave residual features from the original style and neglect content features from style references. We propose SCAdapter, a novel method using the CLIP image space for separating and recombining content and style features. This approach extracts pure content features from the content image and style features from the style reference, ensuring cleaner and more authentic transfers by eliminating cross-contamination. By incorporating this recombined representation into the prompt embedding, SCAdapter provides precise guidance for the diffusion model. Our experiments show SCAdapter’s superior performance, especially in photo-to-photo transfers, confirming its efficacy in style transfer.
Paper :
SCAdapter: A Content-Style Disentanglement Approach for Diffusion-Based Style Transfer
（外部サイト）
PDF : SCAdapter: A Content-Style Disentanglement Approach for Diffusion-Based Style Transfer