Vanast: How AI Is Transforming Virtual Try-On Video for Fashion

Virtual try-on (VTON) has become one of the most heavily funded and active research areas in computer vision, driven by strong investment from the fashion industry. Image-based systems are already commercialized on platforms such as veesual.ai, wanna.fashion, and fashn.ai. Extending this capability from static images to video, however, remains a significant challenge. This article examines Vanast, a new single-stage AI model from Seoul National University that aims to generate stable, realistic virtual try-on videos from a single person photo and a garment image.

Video virtual try-on is far harder than image-based try-on. Systems must preserve the garment's details while the body moves, and they must correctly show the back of a product when the person turns. Earlier efforts, such as Google Labs' experimental Doppl app and academic projects like Fashion-VDM, struggled with low latency and output quality. Doppl was shut down in April 2026 after a lukewarm reception, with users redirected to Google's image try-on service. Most existing video offerings also remain cutting-edge token-based products with no clear link to physical retail.

Vanast combines three inputs, a person photo, a separate garment image, and a motion reference video, into a unified framework. Rather than splitting identity, clothing transfer, and motion into discrete steps, the single-stage design lets these aspects interact and fuse during training, aiming for more coherent results. The system produces a complete video sequence in which the person wears the new garment while following the specified motion, maintaining identity and garment detail consistently across frames.

No public dataset met the project's requirements, so the researchers built a synthetic dataset of triplets (person image, garment image, motion video). Data was collected from online shopping platforms, but because real-world footage rarely shows one person wearing multiple outfits, synthetic generation was necessary. A three-stage pipeline selected suitable candidate frames, processed them through the Qwen2.5-VL vision-language model, and created inpainting masks to isolate affected regions. Qwen was also used for gender classification, while the Flux diffusion framework generated garment modifications. To increase pose and background diversity, the team further extracted training triplets from the HumanVid dataset.

To address slow convergence and weak control balance in earlier methods, Vanast splits its model into two modules: the Human Animation Module (HAM) and the Garment Transfer Module (GTM). Both share access to a backbone and integrate features in a distributed, cascaded manner to improve conditioning. Training freezes the backbone and optimizes only the HAM and GTM parameters, balancing their contributions during feature integration. Inputs are converted into latent representations using the VAE from WAN, motion-aware context is built by combining human and pose information, and garment features are projected into token embeddings. The model can also be extended to support garment interpolation.

The model was trained on 9,135 videos ranging from three to ten seconds, drawn from shopping mall websites, the authors' own generated dataset, and HumanVid. Two evaluation datasets were built: an internet dataset of mall videos and product images, and the official test set of Alibaba's ViViD dataset. Because ViViD lacks faces, faces were added through Flux outpainting. Evaluation metrics included L1 loss, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), Fréchet Inception Distance (FID), and Fréchet Video Distance (FVD).

Vanast was compared against image try-on systems including OOTDiffusion, CatVTON, OmniTry, and Any2AnyTryon, as well as VisualCloze, MOSAIC, UNO, and VACE. In a second category of tests, combinations of image virtual try-on and animation models were evaluated, and Vanast again achieved the highest scores across the reported metrics. These results suggest that the unified approach offers measurable quality gains over modular pipelines.

Despite its promising results, the Vanast paper provides limited detail on training and inference resource requirements, suggesting the system may not be the most flexible or agile option. Virtual try-on remains extremely difficult even in non-optimized systems, and commercial deployment demands low latency and favorable scalability and return on investment. The gap between impressive demonstrations and practical deployment means the field still faces substantial hurdles.

Vanast represents a meaningful step toward coherent, end-to-end video virtual try-on, improving on fragmented pipelines through unified training and synthetic triplet supervision. However, the high computational demands and latency documented in the research mean the technology is not yet commercially practical at scale. Current showcases often mask these difficulties, and the problem may ultimately be solved by lighter, more efficient techniques beyond today's transformer-based approaches.