Voost - An innovative two-way virtual try-on and try-off AI model

What is Voost?

Voost is an innovative virtual try-on and try-off model from NXN Labs, developed based on the unified and scalable Diffusion Transformer (DiT) framework. It can simultaneously handle virtual try-on and try-off tasks, generating high-quality image results. By jointly learning these two tasks, Voost utilizes a bidirectional supervision mechanism, enabling each clothing-person data pair to provide supervisory signals for generation in both directions. This significantly enhances the ability to infer the relationship between clothing and the body, without relying on task-specific networks, auxiliary losses, or additional labels.

Voost's main functions

• Two-way virtual try-on and try-offVoost can handle virtual try-on and try-off tasks simultaneously, generating high-quality image results that allow users to view the effect of wearing and removing the target clothing.
• Unified frameworkBy jointly learning virtual try-on and try-off tasks through a single Diffusion Transformer (DiT), the model structure is simplified and efficiency is improved without relying on task-specific networks, auxiliary losses, or additional labels.
• Enhanced relational reasoningBy utilizing a two-way supervision mechanism, each pair of clothing-person data can provide supervision signals for the generation in both directions, thereby enhancing the reasoning ability regarding the relationship between clothing and the body.
• Robustness improvementAttention temperature scaling technique is introduced to enhance the model's robustness to resolution or mask changes; a self-correcting sampling strategy is adopted, and bidirectional consistency verification is used to improve the stability and accuracy of the generated results.
• High-quality generationIn multiple benchmark tests, Voost achieved the best performance in terms of garment alignment accuracy and visual fidelity, demonstrating excellent generalization ability and generating realistic try-on and try-off images.
• Flexible conditional inputIt supports flexible conditional inputs, allowing conditionalization based on generation direction and clothing category, enhancing the model's flexibility and adaptability, and making it suitable for various clothing categories and human poses.

Voost's technical principles

• Unified Diffusion Transformer FrameworkVoost employs a single Diffusion Transformer (DiT) to jointly learn virtual try-on and try-off tasks. Through a two-way supervision mechanism, each pair of clothing-person data can provide supervision signals for generation in both directions, enhancing the ability to reason about the relationship between clothing and the body.
• Two-way supervision mechanismBy jointly modeling virtual try-on and try-off tasks, Voost utilizes bidirectional supervision signals to improve the model's understanding of the correspondence between clothing and the body, without the need for additional labels or task-specific networks.
• Attention Temperature ScalingThe attention temperature scaling technique is introduced to adjust the attention weights, enhance the robustness of the model to changes in resolution or mask, and ensure stability and consistency under different input conditions.
• Self-correcting sampling strategyCross-consistency verification is performed using bidirectional generation results. A self-correcting sampling strategy is used to improve the stability and accuracy of the generated results, ensuring the visual consistency and realism of the generated images.

Voost's project address

• Project official websitehttps://nxnai.github.io/Voost/
• Github repository: https://github.com/nxnai/Voost
• arXiv technical paper: https://arxiv.org/pdf/2508.04825

Voost application scenarios

• e-commerce platformProvide users with a virtual try-on function to help them see how clothing looks on them more intuitively, improve the shopping experience, reduce the return rate due to unsuitable size or style, and increase the platform's conversion rate.
• Fashion DesignDesigners can use Voost to quickly preview how clothing designs look on different mannequins, assess the feasibility of the designs in advance, optimize the design process, and reduce design costs.
• Personalized customizationIt provides consumers with a personalized virtual try-on experience, allowing them to choose different clothing styles, colors, and combinations according to their needs, thus achieving customized services and meeting individual requirements.
• Clothing ShowBrands and merchants can use Voost to showcase clothing online, attract more users through virtual try-on features, and enhance brand influence and product exposure.
• Virtual fitting roomWe provide virtual fitting solutions for offline clothing stores, reducing customer waiting time, improving fitting efficiency, and providing customers with a richer fitting experience.