Our first purpose-built virtual try-on model. Not a general-purpose diffusion pipeline pointed at fashion - a model, a moderation pipeline, and a routing layer designed around one job: put a real shopper in a real garment, accurately, in seconds.
Pattern, print, and logo preservation across our internal evaluation set of ~12,000 product/pose pairs.
Mean similarity between the shopper's uploaded photo and the generated result - the person stays the person.
End to end, upload to published vibecheck. p95 at 6.1s.
Tops, bottoms, outerwear, dresses, footwear, and accessories, each handled with category-aware fitting.
Caught by the moderation pipeline before generation ever starts.
A routing layer spans multiple try-on engines with automatic failover, so one provider hiccup never becomes a shopper-facing failure.
Internal benchmarks, evaluated against Vibecheck's own generation and moderation logs. Not an independently audited result.
Most try-on demos are a diffusion model given a clever prompt. Vibecheck Model v1 is trained specifically on garment transfer - segmentation-free, so it doesn't need a merchant to pre-mask every product photo, and category-aware, so a flowing dress and a structured jacket are fit differently rather than pasted onto the same silhouette.
Every vibecheck request goes through a routing layer, not a single hardcoded API call. If a provider is slow or degraded, the request fails over to the next one automatically - a shopper never sees a spinner that just gives up, and a merchant never sees a dead Vibecheck button because one upstream provider had a bad day.
Every upload runs through dimension, body-detection, and content-safety checks before a single generation credit is spent - not a manual review queue after the fact. And it's built to be fair, not just strict: signals like religious or cultural coverings are explicitly excluded from anything that would block a shopper's photo.