Solutions
AI-Powered Virtual Try-On
Machine-learning face landmarks keep products aligned naturally as customers move, tilt, and explore — with no cloud AI bill.
AI face tracking for accurate placement
Tryonixs uses Google's MediaPipe Face Landmarker — an open-source (Apache-2.0) machine learning model — to detect 468 facial landmarks and head rotation in real time, directly in the browser. Product assets scale, rotate, and anchor to those landmarks so they follow head pitch, yaw, and roll instead of sitting statically on the video.
For the foundation shade-matching module, Tryonixs adds a second model — BiSeNet, an open-source (MIT-licensed) face-parsing network trained on CelebAMask-HQ — that segments the face into 19 semantic regions (skin, eyes, lips, hair, background) so shade is applied only to skin pixels while texture is preserved.
This is deliberately not a cloud-AI product
A lot of "AI virtual try-on" vendors route your customers' camera frames to a cloud API — OpenAI, a proprietary vision API, or a paid third-party SDK — and bill per inference. Tryonixs's architecture explicitly excludes that model: every AI model it ships is open-source (Apache-2.0 or MIT) and runs entirely on the shopper's device via WebAssembly. There is no per-session AI API cost, and no camera frame ever leaves the browser.
That's a real trade-off, not just a privacy talking point: on-device inference means zero marginal AI cost as you scale to thousands of concurrent shoppers, and it means you're not dependent on a third party's API uptime or pricing changes.
Built for performance at scale
Client-side inference minimizes latency — there's no network round-trip per frame — and keeps your server load flat regardless of how many shoppers are trying products on simultaneously, since the AI workload lives in their browser, not your infrastructure. Models are lazy-loaded and cached in memory per session, and disposed of when the shopper closes the try-on view.
Frequently asked questions
Is camera data sent to your servers?
No. Face tracking and AI inference run client-side in the shopper's browser. Tryonixs only records anonymized analytics events you configure, such as session starts and which product was tried on — never camera frames.
What AI models power Tryonixs?
MediaPipe Face Landmarker (Google, Apache-2.0) for face tracking across all categories, and BiSeNet face parsing (MIT-licensed, ResNet-18 backbone trained on CelebAMask-HQ) for skin segmentation in the foundation shade-matching module — both run via WebAssembly in the browser, not a paid cloud API.
How is this different from AR virtual try-on?
AR virtual try-on describes the visual result — a product overlaid on a live camera feed. AI virtual try-on describes the technology producing that result: machine-learning models detecting facial geometry in real time. Tryonixs uses AI face tracking to power its AR experience — the two terms describe the same product from different angles.
Does on-device AI work on older phones?
Yes, with graceful fallbacks. Tryonixs automatically tries GPU-accelerated inference first and falls back to CPU delegates (and alternate rendering paths on iOS Safari) if the device or browser doesn't support the faster path, so tracking still runs — just at a lower frame budget on older hardware.
Launch virtual try-on today
Free Starter plan — 15 products, unlimited try-on sessions, no credit card required.
Create free account