IDM-VTON and Open-Source Virtual Try-On: What Fashion Brands Should Know Before Shipping

IDM-VTON and other open-source virtual try-on models make convincing on-body images cheap to produce. What fashion brands should know before shipping is that an image is not a sell-in. A try-on picture settles how a garment looks, not whether it fits or can be made, and the disclosure rules may apply to you. Our sources cover no specific try-on model, so we make no claims about IDM-VTON's benchmarks or licence.

Why images are cheap now
More than 18 foundation image models exist. List price per 1024px image fell from about $0.020 (DALL-E 2, 2023) to about $0.0017 (GPT Image 2 standard tier, 2026), about 92 percent. These are list prices, not what brands pay. A picture is no longer proof of anything.

The production gap
In a worked example from Whitepaper 04, a creative director approves a look. The technical designer gets a flat with proportions that do not match, the BOM misses two trims, the vendor sends clarifications and merchandising flags copy that does not match product data. Ten more generated images fix none of it and add ten more things to reconcile.

What a try-on render settles and what it cannot
Question | Settled by a render | Needs production evidence
Silhouette matches the brief | Yes | No
Colorway and line review | Yes | No
Armhole and wearing ease | No | Yes
Seam balance and grade | No | Yes
Fabric stretch and recovery | No | Yes

The image is pixels with no pattern under it. Most disappointment comes from treating visual confidence as production confidence.

Test the claims
Image-derived 3D is a preview for the eye. AI does not replace physical sampling; the honest goal is fewer rounds. A generated tech pack is a draft until a technical designer approves it. This tests claims and ranks no vendor.

Three questions for any vendor
Where does the pattern geometry come from? What fabric data does the simulation run on, and how was it tested? What happens on a style with no history? A tool that answers cleanly produces evidence. One that deflects produces pictures.

Who the disclosure rules bind
Most labeling laws bind whoever builds the model. EU AI Act Article 50 binds providers, with those on the market given until 2 December 2026. California SB 942, as amended by AB 853, binds providers with over one million monthly users, so most brands are very likely not covered. A New York law from 9 June 2026 binds the brand or seller on synthetic performers, with penalties of $1,000 then $5,000. From July 2026 Amazon requires sellers to tag photorealistic AI people with IPTC metadata. A synthetic model in a real garment is in scope. This is context to August 2026, not legal advice.

The F* Word ties images to the tech pack, so what you show matches what the factory makes.