TL;DR:
A virtual try on glasses demo is a best-case recording. I run seven checks on a live camera before I believe one: real-world scale, temple occlusion, head turns, lens rendering, jitter, bad light, and how a frame gets into the catalog. This post is a checklist. I have not published measurements here.
I look at a lot of virtual try on glasses demos for client projects. Most of them look fine in the first ten seconds, because the demo is a frontal face, good light and one hand-tuned frame. Shoppers do not behave like that. They tilt the phone, sit by a window and try thirty frames in a row.
So I test the way a shopper breaks things. Each check below is something you can run yourself in under a minute, on the device you actually care about.
What did I actually measure for this post?
No numbers, and I want that stated before the checks. This is step zero: the method. If I publish measurements later they will be per device, never averaged, with the phone model, OS version and chipset written next to every number. For now, if you run the checks yourself, you will know what to ask any vendor to show you.
How do you test virtual try on glasses on a real face?
Run the seven checks in this order. The early ones fail fastest.
1. Is the frame the right size, in millimeters?
A try-on that looks good but renders every frame at one default width is a styling toy.
Take a physical frame and read the numbers printed on the inner arm, for example 52-18-140. Those are lens width, bridge width and temple length, in millimeters. Note what they are not: 140 is the length of the arm that runs to your ear, not how wide the frame is. The front of that frame measures roughly 2 x 52 + 18 = 122 mm across, plus a millimeter or two of hinge. If a vendor quotes you the third number as a width, that alone tells you something.
Here is the arithmetic you can redo. Say the overlay renders that 122 mm front 2 mm too narrow. That is 2 / 122, about 1.6%. Small on its own, and you probably will not see it.
Now add the pupillary distance. Adult PD runs roughly 54 to 74 mm, with 63 mm a common middle. An estimate that is 3 mm off on a 63 mm PD is a 4.8% error, and it puts each lens center about 1.5 mm away from the eye it is supposed to sit in front of. Combine a 1.6% scale error with a 4.8% PD error and the frame stops sitting where the eyes say it should, even though neither number sounds alarming on its own. Illustrative numbers, so swap in yours.
What I look for: does the vendor use a stated scale reference, and does the same frame in two sizes render at two sizes?
2. Do the temples disappear behind the ears?
Turn your head slowly to the left. On a real face the far temple arm goes behind the ear and the near one stays visible. If the arms float in front of the cheek, the occlusion is missing. This is the single most common tell I see in cheap overlays.
3. What happens at 45 degrees and beyond?
Sweep from frontal to a hard yaw, then pitch up and down. Tracking that holds at 10 degrees and snaps at 40 loses the shopper exactly when they check the side profile. Landmark models tend to be happiest on near-frontal faces, which is why the sweep is worth doing. For reference, Google's MediaPipe Face Landmarker documents 478 landmarks per face, and even a model with that many points can lose them when half the face leaves the frame.
4. Do the lenses look like lenses?
Tinted, mirrored and clear lenses need different treatment. Clear lenses should not turn the eyes into a gray rectangle. Look for glare that moves when you turn, and for frame edges that stay crisp against hair and skin. If every material looks like matte plastic, the asset pipeline is flat.
5. Is the frame steady when you hold still?
Hold the phone against a stack of books and watch for the frame swimming a pixel or two on a face that is not moving. Jitter is cheap to hide in a recording and hard to hide live. I count it as a fail when I notice it without looking for it.
6. Does it survive a window behind you, and your own glasses?
Backlight, a dim room and a face already wearing glasses are three separate tests. Many tracking stacks quietly assume bare faces. If the shopper's own pair confuses the overlay, the first thing they will do is take them off, and many will not.
7. How does a new frame get into the catalog?
This is where projects go over budget, and it is the check almost nobody runs because it is not visual.
Ask four things, in this order:
- What does the merchant supply per SKU? Product photographs, a studio capture session, or a 3D model someone has to build. These are three different budgets.
- How many images per frame, and from which angles? The answer separates a flow you can run on the photos already in your product feed from one that needs a turntable.
- Do colorways need separate capture? A catalog of 400 frames in three colors each is 1,200 SKUs or 400, depending on the answer.
- How long does a collection take to go live, and who did the work in the number you were quoted? "Two days" means something different when the vendor's team did it. A photo-based flow and a 3D-modeling flow have very different costs, and a try-on that works for twenty hero frames and breaks at two thousand SKUs only works as a lookbook.
How do I score what I see?
I use a plain pass / partial / fail per check, written next to the device it ran on. The table is my own scoring sheet. The columns and the thresholds are mine, not an industry standard.
| Check | What I look at | Fail looks like |
|---|---|---|
| 1. Scale | Printed frame size vs on-screen size | Every frame renders the same width |
| 2. Temples | Far arm during a slow head turn | Arms float in front of the cheek |
| 3. Head turns | Yaw and pitch sweep | Overlay snaps or detaches |
| 4. Lenses | Clear, tinted, mirrored | Everything looks matte |
| 5. Jitter | Phone still against a stack of books | Visible swim on a still face |
| 6. Light and own glasses | Backlight, dim room, glasses on | Frame lost or misplaced |
| 7. Onboarding | Images and time per new SKU | Only hero frames are available |
Every row in that table is stated in the text above, so nothing lives only in the table.
Should I test on my own phone or the vendor's?
Yours, plus the oldest mid-range phone in your test pool. A flagship is the best case for every one of these checks, and that is the device the demo was recorded on. I wrote up the same habit for face AR SDKs in 8 Face AR SDK Checks I Run Before Signing a Contract.
Why does color matter for glasses at all?
It matters less than for makeup, but it still shows up in tinted lenses and acetate frames under warm light. The long version of why rendering never matches the real thing is in 6 reasons virtual makeup rendering never matches the real shade.
What this does not cover
This checklist says nothing about purchase conversion, return rates or pricing. It also says nothing about which product to buy, because I have not run these checks against named products for this post. Treat any vendor claim about fit accuracy as a claim until you have run checks 1 to 3 yourself on a few real faces.
If you only run three
Run 2, 3 and 7. Temple occlusion and a hard head turn expose weak tracking in under a minute, and the onboarding question exposes weak economics before you sign anything. Everything else is polish you can negotiate.
Top comments (0)