Content Management · Content & Media
Should you build or buy AI Virtual Try-On (Generative VTON)?
Generative AI virtual try-on (VTON) renders a shopper or model wearing a specific garment from a photo — diffusion/image-model garment transfer, distinct from AR face-tracking try-on (glasses, makeup). Sold as per-image APIs and SaaS by specialists like FASHN AI ($0.075/generation), Botika, Revery.ai, and Vue.ai, as an enterprise API by Google (Vertex AI Virtual Try-On), and increasingly as a prompt-level capability of general image models (Gemini image, GPT-Image). Retailers use it for PDP try-on experiences, fit confidence, and catalog/marketing imagery at scale.
This is a buy — but a short-leash one. The capability is real (documented conversion lifts and return reductions in apparel and footwear), the vendor pricing is already commodity-cheap per image, and the self-build path is weaker than it looks: the leading open models (IDM-VTON, OOTDiffusion, CatVTON) all carry non-commercial licenses, and there is no named evidence of a brand running self-built VTON in production. The reason to stay on short contracts is the absorption trend: Google already folded its dedicated try-on app into its general image model, and general-purpose image APIs now beat specialized VTON models on quality benchmarks. The category you're buying may become a prompt you run against a foundation model API you already pay for.
Build it, buy it, or bridge?
When building makes sense
Almost never, today. The strong open-source models (IDM-VTON, OOTDiffusion, CatVTON) are licensed non-commercial, so a compliant self-build means either negotiating a commercial license or building on general image APIs — which is still buying, just at a different layer. No named brand runs self-built VTON in production, and general-purpose models already top the quality benchmarks, so a dedicated in-house pipeline is a stranded investment the moment the next foundation image model ships. The exception is a marketplace at extreme volume where per-image API spend rivals engineering cost — and even then the build is a routing-and-QA layer over rented models, not a trained try-on model of your own.
When buying makes sense
Buy if you sell apparel, footwear, or accessories and want try-on on product pages or generative catalog imagery: the APIs are commodity-cheap (fractions of a cent to a few cents per image), live in days, and carry documented conversion and return-rate benefits in the categories where fit anxiety is real. Prefer per-image pricing and short terms over long contracts — the capability is deflating and being absorbed by general image models, so the price you'd lock today is likely above the price you'll see next year. Benchmark two or three backends (a specialist like FASHN, Google's Vertex try-on API, and a general image model) on your own catalog before committing; an aggregator that routes between them keeps switching costs near zero.
The desk read
Build-versus-buy analysis for AI Virtual Try-On (Generative VTON) is being written. In the meantime, the framework that drives every B4 call is on the B4 Index page.
Frequently asked
What is generative AI virtual try-on (VTON)?
Image-model garment transfer: a diffusion or foundation image model renders a specific garment on a photo of a person or model. It's distinct from AR try-on, which uses live face/body tracking for glasses, makeup, and accessories.
How is this different from the AR Try-On category?
AR try-on is real-time camera tracking (glasses, makeup, shoes-on-feet); generative VTON is offline image rendering (garment on a photo). They solve different moments — AR for live 'how does it look on me now,' VTON for product pages, fit confidence, and catalog imagery at scale. B4 scores them separately because the build-vs-buy dynamics differ.
Can we self-host an open-source VTON model?
Legally, usually not for commerce: IDM-VTON, OOTDiffusion, CatVTON, and StableVITON all ship under CC BY-NC-SA (non-commercial) licenses. And operationally there's no named evidence of any brand running self-built VTON in production. The practical self-serve path is general image APIs — which is still buying.
Will foundation image models make dedicated VTON vendors obsolete?
The trend points that way: Google shut its dedicated Doppl app and serves try-on from its general Gemini image model, and quality benchmarks now rank general-purpose models above the specialists. Dedicated vendors keep an edge on accessory categories, identity preservation, and catalog-scale QA workflow — for now. Buy short, benchmark often.
What does it cost?
Specialist APIs run about $0.075 per generated image (FASHN, also hosted on fal.ai), with SaaS tiers from ~$19-99/month; Google's Vertex AI try-on API is usage-based; enterprise suites (Vue.ai, Botika) price custom. Consumer-facing try-on in Google Search/Shopping is free.