CFADiff: A Diffusion Model-Based Approach to Cross-Domain Style Clothing Generation

Artificial intelligence-driven digital technologies have profoundly transformed daily life. This paper explores an innovative task in the generation of fashion imagery utilizing advanced AI techniques. Specifically, the goal of this task is to generate novel fashion images by blending texture styles from non-fashion domain images with existing fashion content, offering designers a source of inspiration. Aiming at the problem that the style texture information in style reference images is difficult to define and extract, we proposes CFADiff, an unsupervised cross-domain fashion style clothing generation method based on the diffusion model. The method first uses denoising diffusion implicit model (DDIM) to forward encode the original clothing images to convey structural information, and utilizes deep features extracted by a large-scale self-supervised visual transformer (DINO-VIT) as a priori for style semantics. Finally, the stylistic texture and structural semantics in the reference image are decoupled in an explicit manner for guiding the reverse generation process. Experimental results show that CFADiff performs better in terms of structure preservation, stylistic texture performance, and quality of the generated images compared to the baseline methods compared.