Vanast turns a single photo and outfit into a full try-on video

Virtual try-on tools have been around for a while, but getting an AI-dressed avatar to actually move naturally without morphing into a completely different person has remained notoriously hard.

Researchers unveiled Vanast, a unified framework built to tackle that exact problem. Hand it a single portrait, photos of a garment, and a skeletal pose video, and it synthesizes a coherent animation of the person wearing the new outfit in motion.

Ditching the clunky two-stage pipeline

Most existing solutions try to solve this with a split workflow. First, they apply a virtual try-on model to paste clothes onto a still image. Then, they feed that result into a separate pose-driven video animator.

That handoff rarely works smoothly in practice. Errors compound fast, causing noticeable identity drift in the face, distorted clothing folds, and glaring mismatches when the character turns around.

Vanast replaces that patchwork with a single, end-to-end framework. To train it, the researchers built large-scale synthetic triplet supervision pipelines, pairing identity-consistent people across diverse outfits, full top-and-bottom combinations, and in-the-wild shots.

Under the hood, they introduced a Dual Module architecture for video diffusion transformers. This stabilizes training, locks down the subject's facial identity and body proportions, faithfully reproduces garment details, and even allows for zero-shot garment interpolation.

From tees to dresses, but code is still pending

Visual demos show the model handling standard wardrobe swaps smoothly. Whether putting a new shirt on a guy or fitting a skirt and pants onto a model, both facial features and body geometry stay intact.

It also handles trickier pieces like full-body dresses without tearing or unnatural stretching as the subject moves.

That said, you cannot run it locally just yet. The researchers published the paper, but the official repository lists code as coming soon.

Why it matters

If Vanast delivers on its open-source promise, fashion e-commerce workflows are in for a shakeup. Instead of booking studios and live models for every seasonal catalog drop, brands could generate high-fidelity try-on motion reels directly from flat-lay apparel shots and a few standard movement templates.

On the research side, it sends a clear signal that multi-step try-on pipelines are hitting a dead end. Modeling clothing transfer and temporal motion simultaneously within a unified video diffusion transformer is proving far more robust than chaining isolated 2D tools together.

We think the unified approach is the right engineering call, as eliminating pipeline handoffs directly cuts down on identity distortion. Still, controlled paper demos do not always translate to flawless real-world performance. Once the repo drops, the real test will be how well Vanast handles wild background lighting, extreme poses, and loose, flowing fabrics.