Single-photo virtual try‑on

Online clothing demos usually fake the easy part: one still image. The hard part is motion, because sleeves twist, hems swing, and a jacket has to stay the same jacket when a person turns around. (arxiv.org) Virtual try-on is the field that swaps a new garment onto a person in an image. Older systems often did that in two steps: first make a dressed photo, then animate that photo into video. (arxiv.org) That two-step setup breaks in predictable ways. Researchers call the failures identity drift, garment distortion, and front-back inconsistency, which is a formal way of saying faces change, clothes warp, and the back of a shirt stops matching the front. (arxiv.org) The new paper is called Vanast, from researchers at Seoul National University. It takes one human photo, one or more garment images, and a separate pose-guidance video that acts like a motion script for how the body should move. (arxiv.org) (hyunsoocha.github.io) Instead of handing the job from one model to another, Vanast does the garment swap and the animation in one unified pass. The point of that design is to keep the person’s identity and the clothing details locked together across the whole clip. (arxiv.org) The training trick is what makes that possible. The team says it built large-scale triplet supervision, meaning matched sets of person, garment, and motion examples that teach the model what the same outfit should look like from frame to frame. (arxiv.org) This is different from earlier video try-on work like Fashion-Video Diffusion Model, which starts with a person video. Vanast starts with a single person image, so it asks the model to invent the missing side views and motion while keeping the clothes believable. (dl.acm.org) (arxiv.org) It also lands in a fast-moving corner of research. A 2025 paper called MagicTryOn worked on garment-preserving video try-on, and a 2025 paper from the University of Illinois described arbitrarily long try-on videos, but both still reflect how quickly labs are pushing from static catalog images toward moving previews. (huggingface.co) (yxw.cs.illinois.edu) The commercial pitch is obvious. A retailer that already has flat product shots and one model photo could mock up short moving clips for dozens of outfits before a studio shoot ever happens. (unite.ai) (arxiv.org) The catch is that “looks real” is not the same as “fits right.” Vanast is a computer vision paper about visual coherence, not a sizing system, so it can show drape and motion without telling you whether a medium will pinch at the shoulders. (arxiv.org) The paper went up on arXiv on April 6, 2026, and the project page labels it for Conference on Computer Vision and Pattern Recognition 2026. That means the result is fresh, impressive, and still at the stage where researchers are showing examples rather than retailers reporting conversion numbers. (arxiv.org) (hyunsoocha.github.io)