AnyDesign: Versatile area fashion editing via mask-free diffusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Image Processing

Background:

• Current fashion image editing methods lack flexibility and a unified framework, often requiring auxiliary tools like segmenters and keypoint extractors.
• Existing datasets and methods are limited to generic garments and simple backgrounds, restricting real-world applicability.
• The need for versatile and user-friendly fashion editing tools is significant for both research and practical applications.

Purpose of the Study:

• To develop a mask-free, flexible, and unified framework for fashion image editing.
• To enhance datasets with diverse apparel and complex backgrounds for more robust training.
• To enable users to edit fashion images using text or image prompts without manual segmentation.

Main Methods:

• Extended an existing human generation dataset to include a wider variety of apparel (e.g., headwear, scarves, bags) and complex backgrounds.
• Proposed AnyDesign, a diffusion-based method for mask-free fashion image editing.
• Incorporated Fashion DiT with a Fashion-Guidance Attention (FGA) module to fuse apparel type and CLIP-encoded features.

Main Results:

• The extended dataset provides a richer resource for training versatile fashion editing models.
• AnyDesign successfully performs mask-free editing on diverse clothing items and complex scenes.
• Qualitative and quantitative experiments show AnyDesign produces high-quality results and outperforms existing text-guided methods.

Conclusions:

• AnyDesign offers a significant advancement in fashion image editing by providing a flexible, unified, and mask-free approach.
• The method demonstrates superior performance in handling diverse fashion items and complex scenarios.
• This work paves the way for more realistic and accessible AI-powered fashion content creation.