
In 205,000 Generated Try-Ons, Drape Passed on Looks and No Garment Had a Size
CURVTON-205K, a research dataset of try-on triplets, was drawn almost entirely by FLUX.2 [klein] 9B and screened by a vision-language model and a filter trained on a ten-question rubric that asks whether fit and drape look plausible. Its people carry BMI bands and its garments no size label, so it can teach a believable drape, not a fit.
Parallax Pincer
A white tunic lies flat on grey concrete, long-sleeved, with blue geometric embroidery down the placket and round both cuffs. In three of the four try-ons that put it on a long-haired blond man kneeling on a beach, the sleeves are gone and the cuffs sit on bare wrists or round an ankle. The garment is a Ghalek, and a new paper’s Figure 13 shows the strip as part of the retry loop behind CURVTON-205K. That dataset is meant to teach try-on models how clothes sit on bodies.
The words and the picture disagreed. The project’s garment dictionary lists the Ghalek as a “sleeveless overdress falling to ankle, worn over a blouse”. The released code turns such entries into the generator’s first instruction, after “Make this person wear this dress”. Three renderings drop the sleeves, as the sentence does; the fourth keeps them, as the picture does. The caption calls the triplets accepted and lists no artifacts; it does not mark which of the four the checker accepted.
CURVTON-205K, posted on 18 November and accepted at WACV 2027, comes from researchers at IOE Pulchowk Campus (Tribhuvan University) and NAAMII in Nepal and at Fogsphere in Britain. It holds 205,000 training and 4,500 test triplets, each a person, a garment and the person wearing it, in easy, medium and hard tiers. FLUX.2 [klein] 9B drew 35,000 garments, 41,000 people, 300,000 edited versions of them and every try-on. The released code also moves the camera on 15% of medium and hard triplets, redrawing person and try-on with Qwen Image Edit, much as Phota re-angles a snapshot; the paper describes camera changes only as FLUX edits.
The variety is written into prompts. People are sampled from an attribute dictionary of body types, ethnic groups, backgrounds and photo styles, and one in five draws a visible disability, from a missing finger to a wheelchair. Half are prompted as studio shots and half as phone snapshots. The paper counts 200 garment types, many traditional, and the appendix shows a qipao, a gomesi and a Romanian ie.
Fit is judged by eye. Qwen3-VL 32B compared each try-on with its person and garment and returned failures with a rewritten prompt, up to four times. Its instructions give it the garment’s description, “sleeveless” included, and tell it to check “sleeve shape and length” against the photo, while allowing “natural changes in folds, wrinkles, perspective, shading, and fabric draping caused by the person’s body shape and pose”. The paper reports training-set artifact rates falling from 38%, 48% and 60% (easy, medium, hard) to 3%, 4% and 5% after the fourth pass.
Annotators labelled 5,000 triplets on a ten-question rubric, and a small filter trained on their labels screened the rest. One question asks whether folds, fit and drape “respond naturally to gravity and the person’s pose”, another whether the whole looks “visually and physically plausible”. Only a “severe” mark made a label negative, so a drape marked minor still passed the labels; the filter then kept only scores of 0.85 or higher.
No garment in the dictionaries, prompts and scripts the repository publishes is given a size label.
Bodies get numbers, of a sort. The person dictionary crosses short, average and tall with three BMI bands from 14 to 30, so heights are words and weights are BMI ranges. An easy-tier edit can make a person “plus size”, a “full heavy-set generous frame”. Garment descriptions carry fit words such as “fitted”, “loose” and “form-fitting”, and hems are placed by the body, “knee-length” or “falling to mid-thigh”; none holds a measurement.
Synthetic data can carry a tape measure. FIT, a SIGGRAPH 2026 dataset, drapes generated garments “via physics simulation” and pairs more than a million triplets with “precise body and garment measurements”. Its authors write that current try-on methods “default to generating well-fitted results regardless of the garment or person size”. CURVTON’s try-on prompt carries a garment description and no measurement.
On the paper’s own tests, the data helps. Added to real training sets, it improved most cross-benchmark scores for CatVTON, IDM-VTON, StableVITON and OOTDiffusion, and cut FID, a distance from real photos, by 4.86 to 8.92 points on StreetTryOn, where the baseline is unpaired samples from different real datasets. On VITON-HD and DressCode, training on CURVTON alone trails real-only training on every reported metric, and adding it gains 0.11 to 0.55 FID. Twenty annotators, marking with the same ten questions, gave OOTDiffusion trained on CURVTON 35% of quality votes, against 15% for the next model.
The limitations section allows that the generated world may not fully capture “garment materials” and “person–garment interactions”, which are what fit is made of. The paper never tests fit, so this part is inference: a model trained this way is taught to hang a garment believably on a body, with no size to get wrong. The licence keeps that lesson out of commercial try-on, at least for now: CC BY-NC 4.0, plus FLUX.2 [klein] 9B’s non-commercial terms.