AnyStyle: A Single LoRA is Sufficient for Image-Guided Style Transfer
2026-07-06 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors address the problem of applying the look of one image's art style to another image while keeping the original picture's structure intact. They found that using separate tools to handle style and content often causes conflicts and makes it hard to balance between them. Instead, they show that using guidance directly from the content image, without extra training, works better and faster. Their new method, AnyStyle, uses one unified tool to capture style and this training-free guidance to keep structure, leading to better and more stable results. They tested their approach widely and showed it does well both in quality and accuracy.
image-guided style transfersemantic structurediffusion modelsadaptersLoRAcontent-style disentanglementstructural guidanceperceptual qualitytraining-free methods
Authors
Yongwen Lai, Chaoqun Wang
Abstract
Image-guided style transfer aims to apply the artistic characteristics of a style image to a content image while preserving its semantic structure and layout. Despite advances in diffusion-based methods, existing approaches often face challenges in disentangling content and style, particularly when independently optimized adapters are naively combined, causing conflicts between adapters and limiting controllability over the content-style balance in inference. We further demonstrate that training-free structural guidance directly derived from the content image through the internal attention of pre-trained model outperforms a dedicated content LoRA adapter in terms of structural fidelity and computational efficiency. Building on these observations, we propose AnyStyle, a streamlined framework for image-guided style transfer. The framework adopts a unified single-adapter paradigm for coherent style capture from the style image and incorporates training-free structural guidance from the content image, thus avoiding complex entanglement between multiple adapters and improving controllability and stability. Extensive experiments show that our method delivers competitive quantitative performance and significantly improved perceptual quality. Code is available at https://github.com/Yvan1001/AnyStyle.