The rendered jacket holds its weight. Lapels behave the way wool flannel behaves under studio light, with a slight gravitational sag at the front edge and a half-shadow inside the lining. This is from Tstars-Tryon 1.0, the virtual try-on system Taobao deployed at industrial scale before publishing the paper this April. The drape problem that haunted try-on demos through 2024 is no longer the interesting problem; silk reads like silk now, knit reads like knit.

The paper claims robustness across eight fashion categories, multi-image composition up to six reference images, and stable output through extreme poses and illumination shifts. What it does not claim, because it does not need to, is that the catalog feeding it is structured. The system runs at industrial scale on the Taobao app, serving millions of users and tens of millions of requests, and Taobao supplies its own catalog. That is the input most retailers cannot replicate.

The thing they did not have to invent was the garment data.

Outside the Alibaba estate, virtual try-on demos still rely on hand-curated reference shoots: six clean images per SKU, neutral background, consistent lighting. A separate April paper on fit-aware try-on names the gap in research terms: there are no training datasets that provide precise garment and body size information for ill-fit cases, because nobody photographs garments that look wrong. The model cannot render a poor fit if no one has ever labelled one.

This is the Cardin problem in new clothes. When Pierre Cardin scaled his couture name into licensed ready-to-wear through the 1960s, the silhouettes survived translation; the specifications did not. Within two decades the name covered hundreds of licensees across continents, and the brand experience fragmented with them. Try-on has the same shape: the rendering engine is the silhouette, the catalog metadata is the spec sheet. A retailer pointing Tstars-Tryon at a database where sleeve length reads “regular” will get back a regular sleeve, plausibly rendered, on the wrong body.

The venue of the bottleneck has moved. For three years the trade press treated virtual try-on as a model problem — too synthetic, too uncanny, too slow. The Taobao deployment closes that chapter. The next twelve months sort retailers by who has structured fabric weight, garment length, fit grade, and material composition in queryable form, and who has those values living in PDF lookbooks and spreadsheets named final_v3.xlsx. The latter group will buy the model and find out it does nothing for them.

The aesthetic consequence is that the look of e-commerce is going to fork. Catalogs with clean metadata will publish garments that move on the body. Catalogs without will publish ghosts: flat overlays that betray the rendering engine inside two seconds of scroll. The technology stopped being the variable in April. The catalog is.