This paper proposes a novel multimodal fashion image editing architecture, the Detail-Preserving Diffusion Model (DPDEdit). Based on a latent diffusion model, DPDEdit integrates text prompts, region masks, human pose images, and clothing texture images to guide fashion image generation. Grounded-SAM is used to predict the editing region based on the user's textual descriptions, and local editing is performed by combining these with other criteria. To transfer the details of a given clothing texture to the target fashion image, we propose a texture infusion and enhancement mechanism that preserves high-frequency details of the generated clothing texture using a separate cross-attention layer and an auxiliary U-Net. Furthermore, we extend the VITON-HD dataset with a multimodal large-scale language model to generate paired samples of texture images and textual descriptions. Experimental results demonstrate that DPDEdit outperforms state-of-the-art methods in terms of image fidelity and consistency with the given multimodal input.
