Skip to main content
    Back to BlogTechnology

    Virtual Watch Try-On: How We Built Our Own Wrist Model

    June 29, 202610 min read
    Virtual Watch Try-On: How We Built Our Own Wrist Model

    Eyewear try-on is a mature technology. Point a camera at a face, the glasses appear, turn your head and they follow. Watch brands look at that and ask for the same thing on a wrist.

    It isn't the same thing. A wrist is a much harder target than a face. The watch also has to be shown at its true size, because case size is what watch shoppers worry about most. We found that generic hand-tracking models don't answer either problem well enough, so we built our own wrist model. This post explains what makes watch try-on hard, what we built, and what to check when you compare vendors.

    The Short Version

    • Watch try-on is real-time AR. The watch renders live on the shopper's camera feed and follows the wrist as it turns. It isn't pasted onto a still photo.
    • Generic hand trackers find the hand, not the watch position. They locate knuckles and the wrist joint. They don't say where a watch face sits or how wide the wrist is at the strap line.
    • TryOn Virtual trained its own wrist model to fill that gap. It adds 4 watch-specific points to each hand: the watch-face center, its orientation, and both edges of the wrist where the strap sits.
    • Size comes from the wrist itself. Because the model finds both edges of the wrist, it measures wrist width directly and scales the watch to it. A 38 mm case and a 44 mm case look as different as they do in real life.
    • It runs on the shopper's device. The model is about 4 MB and runs in the browser. Camera frames never leave the phone.

    You can try it live on your own wrist before reading further.

    Why a Wrist Is Harder Than a Face

    Eyewear try-on uses decades of face-tracking research. A face is large and high-contrast, and it nearly always faces the camera. Eyes, nose, brow and jaw give a tracker hundreds of stable points; MediaPipe's face mesh tracks 478 of them.

    A wrist offers far less to work with:

    • Few distinctive features. A wrist is a smooth, roughly cylindrical surface, with nothing like eyes or a nose to lock onto.
    • Fast rotation. People flick and turn their wrist much more freely than their head, and the watch has to stay on the right spot through all of it.
    • Self-occlusion. The hand, a sleeve or the wrist itself hides part of the strap line at most angles.

    This is why watch try-on arrived years after eyewear try-on, and why many watch try-on demos slide, float or change size from frame to frame.

    What Off-the-Shelf Hand Tracking Misses

    The standard starting point is Google's MediaPipe Hand Landmarker. It finds 21 points on a hand: the wrist joint, the knuckles and the finger joints. It is fast and robust, and we use it too.

    Those 21 points were designed for gestures, not for wearing a watch:

    • There is no "watch center" point. The closest landmark is the wrist joint, but a watch sits a few centimetres up the forearm from it. The exact spot shifts with how the hand is held.
    • There is no wrist width. You can estimate width from the knuckles, but palm width is not wrist width. In our tests that estimate was the noisiest signal in the whole pipeline, and wrist width decides the watch's size.
    • Google doesn't release MediaPipe's training code. You can't retrain the hand model to add the points you need. You can only build on top of it.

    So we built on top of it.

    How Our Wrist Model Works

    We call it WristAnchorNet. It runs as a second stage right after hand detection:

    1. Find the hand. MediaPipe locates the hand and its 21 landmarks.
    2. Crop the wrist. We cut a square region around the wrist, shifted toward the forearm. The model then only looks at the area that matters.
    3. Predict the watch points. Our model, a compact MobileNetV3 network, predicts 4 points inside that crop:
      • Watch center: where the watch face sits.
      • Direction: which way the watch points along the forearm, which sets its rotation.
      • Radial edge: the thumb side of the wrist at the strap line.
      • Ulnar edge: the pinky side of the wrist at the strap line.
    4. Place and size the watch. The two edge points give the wrist's real width at the strap line, and the watch is scaled to it. The center and direction set its position and rotation.

    TryOn Virtual's wrist model adds four watch points to standard hand tracking: the watch center, its direction toward the fingers, and the radial and ulnar edges of the wrist, which give the wrist width the watch is sized to

    The result is 25 points per hand: Google's 21 plus our 4. The radial and ulnar edges are named after the forearm bones, the radius and ulna, so they keep the same meaning whichever way the hand faces. That keeps the model consistent for left and right hands and for palm-up and palm-down poses.

    How we trained it

    • Our own labelled data. We built an in-house annotation tool and hand-labelled about 1,700 wrist photos, covering left and right hands, palm and back, and different angles. The tool snaps the two edge points perpendicular to the forearm, so labels stay consistent between people.
    • A 3D hand model to speed up labelling. For the strap-edge points, a 3D hand-mesh model proposes a first guess, and a person corrects it. Correcting a guess is much faster than placing every point from scratch.
    • Training for stability. A try-on that is right on average but shakes from frame to frame still looks broken. We train with a consistency step that penalizes the model for answering differently on two slightly shifted copies of the same image. Every new model is also scored on how steady it stays when the same photo is nudged, rotated and relit.

    How accurate it is

    On wrist images the model never saw during training, the shipped model's median orientation error is 2.5 degrees. In practice, the watch face lines up with the forearm the way a real watch does. We keep training new versions and measuring them against the same held-out set.

    Why it runs in the browser

    The exported model is about 4 MB. It loads with the try-on and runs on the shopper's own device through ONNX Runtime Web. That has three benefits:

    • Privacy. Live camera frames never leave the phone. Nothing is uploaded.
    • Speed. There is no network round trip per frame, so the watch keeps up with the wrist.
    • No app. Shoppers tap a button on the product page. There is nothing to download.

    Case Size Is the Whole Point

    With glasses, fit is mostly about style. With watches, case diameter is a number shoppers worry about and often get wrong. A 36 mm dress watch and a 44 mm diver are different products on the wrist. "It was bigger than I expected" is one of the most common reasons watches come back. Catalog photos on a white background give no sense of scale.

    Because our model measures the shopper's actual wrist width, the watch is drawn at its real proportion to that wrist. A shopper can see that a 42 mm case covers most of a slim wrist while a 38 mm sits more modestly. That answers the question a product photo can't.

    A 38mm and a 42mm watch case shown at true proportion on the same wrist, illustrating the size difference

    Materials: Making Steel Look Like Steel

    Tracking gets the watch in the right place; materials make it look like the real thing. A watch combines several of the hardest materials to render:

    • Polished and brushed metals, like steel, gold and two-tone, which have to catch and move light as the wrist turns. Gold that looks flat or grey undersells the watch immediately.
    • The dial and crystal, a glossy surface with its own reflections under glass.
    • Straps and bracelets, whether link bracelets, grained leather, rubber or ceramic, each needing its own surface finish.

    TryOn Virtual renders watches with physically based materials and HDR environment lighting, so metal reflects a real-looking environment as the wrist rotates. For a premium watch this matters: a cheap-looking render is worse than no try-on at all.

    Eyewear vs Watch Try-On, Side by Side

    Eyewear try-onWatch try-on
    Tracking targetFace (478 points, feature-rich)Wrist (few features, fast rotation)
    What TryOn Virtual usesFace mesh trackingMediaPipe hand landmarks + our own wrist model (25 points per hand)
    Why shoppers hesitateDoes this style suit my face?Does this size suit my wrist?
    How size is setFace widthMeasured wrist width at the strap line
    Hardest materialsLens tint and transparencyPolished metal, gold, dial reflections
    Where it runsIn the browser, on the deviceIn the browser, on the device
    StatusLiveLive

    What to Test When You Compare Vendors

    Any try-on can put a watch on a wrist in a demo video. Test these five things on your own phone:

    1. Turn your wrist quickly. Does the watch stay on the strap line, or slide toward the hand?
    2. Compare two case sizes. Does a 38 mm and a 44 mm look different on your wrist, by the right amount?
    3. Hold still. Does the watch sit still, or shimmer and change size?
    4. Look at the metal. Does gold read as gold under your room's light?
    5. Check where the camera feed goes. Is processing done on the device, or are frames uploaded?

    TryOn Virtual was built around those five tests.

    Frequently Asked Questions

    Is watch virtual try-on real AR or a photo overlay? Real-time AR. The 3D watch renders live on the shopper's camera feed and follows the wrist as it moves and rotates. Clothing and shoes use AI photo try-on; watches and glasses use live tracking.

    Did TryOn Virtual build its own wrist tracking? Yes. Hand detection uses Google's MediaPipe, and on top of it runs WristAnchorNet, our own model. We trained it on about 1,700 hand-labelled wrist photos. It adds four points that standard hand trackers don't provide: the watch-face center, the watch direction, and both edges of the wrist at the strap line.

    Can customers see the actual case size on their wrist? Yes. The model measures the shopper's wrist width from the two strap-line edge points and draws the watch in proportion to it. A shopper can see how a 38 mm and a 42 mm case sit on their own wrist, and misjudged case size is one of the biggest causes of watch returns.

    Does the camera feed get uploaded? No. The wrist model is about 4 MB and runs in the shopper's browser. Live camera frames stay on the device.

    Will a luxury watch look cheap in try-on? Not when the model and materials are good. TryOn Virtual uses physically based materials and HDR lighting, so polished gold, brushed steel and glossy dials reflect light convincingly as the wrist turns.

    Does it work on Shopify? Yes. Install TryOn Virtual on Shopify and a try-on button appears on your product pages. WooCommerce and custom storefronts use a script embed. Shoppers don't download anything. Watch AR is available on every plan, including the free plan.

    Can the same model handle bracelets? The two strap-line edge points are exactly where a bracelet or cuff sits, so bracelet AR is built on the same wrist model. It is currently in beta. See our jewelry try-on guide.

    The Bottom Line

    From the outside, watch try-on looks like eyewear try-on on a wrist. Underneath, it is a harder tracking problem, and it has to answer a question eyewear never does: how big is this on me? That's why TryOn Virtual built its own wrist model rather than relying on generic hand landmarks. It finds where a watch actually sits, measures the wrist it sits on and runs on the shopper's phone.

    Try the live wrist demo, or start free on Shopify and put it on your own catalog.


    This article is part of The Complete Guide to Virtual Try-On in Ecommerce, our full breakdown of try-on types, product categories, ROI and platform options.

    We use cookies to measure how the site is used and how visitors find us. Essential cookies are always on. See our Privacy Policy.