AnyDressing: Customizable Multi-Garment Virtual Try-On via Latent Diffusion Models 208

AnyDressing is a customizable multi-garment virtual try-on method based on a latent diffusion model, proposed by ByteDance and Tsinghua University. AnyDressing consists of two main networks, GarmentsNet and DressingNet, which are responsible for extracting detailed garment features and achieving multi-garment virtual try-on, respectively. AnyDressing can be applied to a variety of virtual try-on scenarios, including both realistic and stylized scenes, providing users with a customized virtual try-on experience.

Main Features of AnyDressing:

Multi-Garment Combination Try-On: Users can try on multiple garments simultaneously, and the system is capable of handling complex garment combinations, ensuring that the details and fit of each garment are well presented.

Personalized Customization: Users can customize try-on effects according to their own needs by providing reference garments and text prompts. This personalized feature allows users to make choices based on their own style and preferences.

High-Quality Image Generation: AnyDressing leverages advanced image generation technology to produce realistic virtual try-on images, helping users better understand how the garments will look on themselves.

Wide Range of Applicable Scenarios: This technology can be applied to multiple fields such as fashion retail, online shopping, and social media, enhancing the shopping experience and satisfaction of users.

Research and Development: AnyDressing is currently in the research stage, jointly developed by the research teams of ByteDance and Tsinghua University. It is mainly intended for academic research and effect demonstration, and has not yet been officially commercialized.

Technical Architecture of AnyDressing:

The technical architecture of AnyDressing consists of two main networks: GarmentsNet and DressingNet. GarmentsNet is responsible for extracting detailed garment features, while DressingNet generates customized images. The garment feature extractor introduced in GarmentsNet is capable of encoding textures of multiple garments in parallel, avoiding garment confusion while ensuring network efficiency. DressingNet adopts an adaptive attention mechanism and an instance-level garment positioning learning strategy to accurately inject multi-garment features into corresponding regions, thereby improving the diversity and controllability of the generated images. In addition, AnyDressing introduces an enhanced texture learning strategy to improve the detail performance of garments. Experimental results of this technology show that AnyDressing achieves state-of-the-art results in multi-garment virtual try-on tasks, and has high practical value and research potential.

The design of AnyDressing addresses the challenges faced by existing technologies in multi-garment synthesis, such as garment fidelity, text consistency, and compatibility with community control extensions. Through its adaptive Dressing-Attention mechanism and instance-level garment positioning learning strategy, it precisely incorporates garment features into the correct positions, while enhancing detail quality through a garment-enhanced texture learning strategy, ensuring that the generated virtual garments have a high degree of realism.

Related Links for AnyDressing:

Project Website: https://crayon-shinchan.github.io/AnyDressing/

HuggingFace: https://huggingface.co/papers/2412.04146

Code: https://github.com/Crayon-Shinchan/AnyDressing

Paper: https://arxiv.org/abs/2412.04146