Voost: A Virtual Try-On and Try-Off Tool 85

Voost is a virtual try-on and try-off AI tool proposed by Seungyong Lee and Jeong-gi Kwak. Its core idea is to handle both try-on and try-off tasks simultaneously using a Diffusion Transformer (DiT). What makes it stand out is that it doesn't require separate task-specific networks or additional labels, and it enables bidirectional "supervised" learning between each outfit and the person — looking at both the try-on result and the try-off result — so it can more accurately understand the relationship between clothing and the body, such as how fabric conforms to the skin and how wrinkles form. It also incorporates two practical techniques: attention temperature scaling (to make the model more stable across different resolutions) and self-corrective sampling (to reduce errors by having try-on and try-off verify each other). In tests, it achieves top-tier performance on both try-on and try-off benchmarks, with impressive reproduction of clothing details (like texture and wrinkles) and good fit with various body poses. It also adapts well to different clothing types, materials, and poses.

📚 Features

• Bidirectional learning for detail restoration: Try-on and try-off are learned together, and through bidirectional supervision the model better understands the interaction between clothing and the body, reproducing details like wrinkles and fabric texture more realistically.

• Realistic generation results: It can display try-on effects from multiple angles, and regardless of body type or pose — standing, sitting, or crouching — the clothing looks more true to life.

• Strong adaptability with minimal dependencies: No need to specially tune the model for a specific garment or scenario, and no need for large amounts of manual labeling. Ordinary photos or webcam captures work fine.

📚 Technical Principles

• Core method: A Diffusion Transformer (DiT) serves as the "brain." Through cross-attention, garment features and body representations are "fused" together, allowing even fine details like a single thread or a single wrinkle to be synthesized precisely.

• Optimization techniques:

• Attention temperature scaling: Like tuning a "sensitivity switch," it keeps the model's output stable under different resolutions or occlusions (for example, when a hand partially covers the clothing).

• Self-corrective sampling: Try-on and try-off "check" each other — for instance, the try-off reverse logic is used to verify the try-on result, reducing generation errors.

• Operation flow: The user uploads a clear front/side photo of a person, or simply takes one with a webcam. The system automatically builds a body model and "puts" the clothing on it.

📚 Where can Voost be used?

• Online clothing shopping: No need to try things on physically — see directly in the app how an outfit looks on you, avoiding wrong sizes or bad fits after purchase.

• Fashion design: When designers tweak silhouettes or mix and match outfits, they can preview results virtually instead of making sample garments over and over.

• Social sharing: Use Voost to "try on" clothes in social platforms, making outfit posts more fun and engaging.

• Offline retail stores: Combined with AR technology to create a virtual fitting mirror — customers stand in front of it and instantly switch outfits, delivering a high-tech experience.

📚 How to use Voost — Step by Step

The Hugging Face public demo is coming soon.

1. Choose try-on or try-off mode;

2. Choose the desired garment type (try-on includes upper body, lower body, and full outfit);

3. Upload a model image (optional: use a drawing tool to create a mask);

4. Upload a garment image;

5. Click the button to run the model.

📚 Frequently Asked Questions

❓: Compared with other try-on tools, what is Voost's biggest advantage?

A: It uses a single Diffusion Transformer to learn try-on and try-off at the same time, without separate task-specific networks or extra labels, and it uses bidirectional supervised learning. Combined with attention temperature scaling and self-corrective sampling, the results are more stable and more realistic.

❓: In which tests has Voost been validated?

A: It has been tested on common datasets such as VITON-HD and DressCode, as well as on real-world photos. Both try-on and try-off results are currently state-of-the-art, outperforming tools such as StableVITON and IDM-VTON in accuracy and realism.

❓: What do attention temperature scaling and self-corrective sampling do?

A: Temperature scaling keeps the model stable under different resolutions or occlusions (for instance, when a hand covers part of the clothing). Self-corrective sampling uses try-on and try-off to verify each other, reducing generation errors and making clothing appear more natural.

📚 Github: https://github.com/nxnai/Voost