The year/Independent research

Paper 2508.04825

Voost: A Unified and Scalable Diffusion Transformer for Bidirectional Virtual Try-On and Try-Off

Published
Aug 2025
Research lab
Independent
Citations
12
GitHub
340 stars

01 In brief

Summary

Voost is a unified and scalable diffusion transformer framework that jointly learns virtual try-on and try-off using a single model.

It uses a token-level concatenation of garment and person images, with a task token encoding generation direction and garment category, enabling bidirectional supervision without task-specific networks or auxiliary losses.

The model supports dynamic layouts and variable aspect ratios.

Two inference-time techniques are introduced: attention temperature scaling for robustness to resolution or mask variation, and self-corrective sampling that leverages bidirectional consistency between tasks.

Experiments on VITON-HD and DressCode show state-of-the-art results, outperforming baselines like StableVITON, OOTDiffusion, IDM-VTON, CatVTON, and Leffa in metrics such as SSIM, LPIPS, FID, and KID.

For try-off, Voost significantly improves FID and KID over TryOffDiff and TryOffAnyOne.

Ablations confirm the benefits of dual-task training, attention-only fine-tuning, and the inference-time refinements.

User studies show Voost is preferred for photorealism, garment detail, and structure.

Limitations include lack of explicit control over garment fit, with future work planned on incorporating body measurements and garment metadata.

02 From the paper

Abstract

Virtual try-on aims to synthesize a realistic image of a person wearing a target garment, but accurately modeling garment-body correspondence remains a persistent challenge, especially under pose and appearance variation. In this paper, we propose Voost - a unified and scalable framework that jointly learns virtual try-on and try-off with a single diffusion transformer. By modeling both tasks jointly, Voost enables each garment-person pair to supervise both directions and supports flexible conditioning over generation direction and garment category, enhancing garment-body relational reasoning without task-specific networks, auxiliary losses, or additional labels. In addition, we introduce two inference-time techniques: attention temperature scaling for robustness to resolution or mask variation, and self-corrective sampling that leverages bidirectional consistency between tasks. Extensive experiments demonstrate that Voost achieves state-of-the-art results on both try-on and try-off benchmarks, consistently outperforming strong baselines in alignment accuracy, visual fidelity, and generalization.