# Virtual Try-On Latency and Mobile Conversion

Mobile try-on fails in the wait, not the idea. Map what shoppers feel, set a budget, show progress and a retry, then measure who leaves before the image arrives.

**Virtual try-on on a phone has to feel fast enough that the shopper is still on the product page when the image arrives. Map the wait they actually feel, from tap to first feedback to finished preview, and design for the long middle. Set a budget, show progress and a way to retry, then measure who leaves before the result. A generation that finishes after they have gone back to Instagram did not convert anyone.**

![Person on worn stoop steps holding a phone with both hands, solid phone back toward us, Antla on a step patch](/images/blog/merchant-100/virtual-try-on-latency-mobile-conversion.webp)

*The wait is the product, from the shopper's side of the phone. Editorial image in Classic Antla disposable-camera style.*

This is not a comparison of your app and your mobile site. That choice has its own tradeoffs, and [mobile app versus mobile web try-on](https://antla.io/blog/mobile-app-vs-mobile-web-virtual-try-on-fashion) already walks them. This is the seconds between "try it on" and a picture, on whatever surface you shipped, and what those seconds do to add-to-cart.

Merchants ask for a single number: how fast is fast enough? The honest answer is a budget with a feedback rule, not a slogan. Human attention has been studied for decades. Your generation time either respects it or it does not.

## Map perceived wait

Jakob Nielsen's [response-time limits](https://www.nngroup.com/articles/response-times-3-important-limits/) are still the cleanest map I know, because they describe what a person feels, not what a server log calls success.

- **About 0.1 second:** it feels immediate. No extra feedback required.
- **About 1 second:** they notice the delay. Their flow of thought is still intact. You do not need a production number, but you have left "instant."
- **About 10 seconds:** that is the edge of keeping attention on the task. Past that, people want to do something else, and they need feedback that says the system is working and roughly how long it will take. For waits longer than about 10 seconds, Nielsen argues for a percent-done indicator, and a way to interrupt.

A try-on preview is not a column sort. It will not hit 0.1 seconds, and it often will not hit 1 second. The design problem is the band between 1 second and well past 10.

Map the wait the shopper feels, in order. Server time alone flatters you.

| Moment | What they experience | What to log |
|---|---|---|
| Tap "try on" | Did anything change in the first tenth of a second? | `modal_open` time |
| Terms, if you show them | Reading is not latency. A frozen screen is | Time until the dialog is usable |
| Pick or shoot a photo | Their time, not yours. Do not bury it inside "generation" | Upload start |
| Upload | A bar that lies at 100% and then sits there | Upload success or error |
| Generation | The long middle. This is where they leave | Generation success or error, with duration |
| Result on screen | First pixel of the preview, not "job queued" | Time to image painted |

Antla lets merchants switch Fast and Pro from the dashboard. Fast is the quicker path. Pro is the higher-fidelity path. [How long virtual try-on takes](https://antla.io/blog/how-long-does-virtual-try-on-take) is the dated public record: Antla does not publish a median generation time, and listing seconds from other vendors are not your catalog. Time Fast and Pro on a phone. Both paths usually sit at or beyond Nielsen's 10-second attention edge, which is why a spinner with no words is the wrong interface for either one.

Add the phone. A generation that already sits past ten seconds plus a slow upload on a 4G handoff is the wait they feel. Measure it on one mid-range phone you keep in a drawer, not only on the office Wi-Fi. [How the technology behaves](https://antla.io/blog/virtual-try-on-technology-fashion) explains why the image takes work. The shopper does not want that essay in the modal. They want to know it is still going.

Perceived wait also includes doubt. If the first second shows no change, they tap again, and you now have two jobs and a double-charged feeling even if you did not double-charge them. The first tenth of a second has to acknowledge the tap. That acknowledgement is cheap. The preview is not.

## Set performance budget

Write the budget down before you argue about models. A budget is a set of ceilings you are willing to ship, with a fallback when you miss them.

I would start from the feeling map, not from a competitor's screenshot:

1. **Tap acknowledgement:** under a tenth of a second, a visible press state. This is interface work, not model work.
2. **Dialog usable:** under a second, including the terms line if you show one. Do not block the dialog on a tracking script.
3. **Generation:** tell the truth about the path you enabled. Time Fast and Pro on a mid-range phone, then put that measured wait in the status line. Do not design the modal as if the image were already in the cache, and do not print a vendor second you did not time.
4. **Failure:** an error by 40 seconds is kinder than a spinner at 90. Pick a timeout and show a retry. An infinite spinner is a crash that still looks busy.
5. **Repeat view:** a shopper who already uploaded in this browser should not repeat the pick-a-file step if the cookie still has their photo. That saved step is part of the budget. The generation may still take the full time. Per [Antla's privacy policy](https://antla.io/privacy-policy): Images are deleted and not accessible within 72 hours. Cookie reuse is a browser convenience, not that deletion clock.

The budget is not "as fast as possible." "As fast as possible" lets a slower Pro default hide inside a button that looks like a filter. Choose the path on purpose. Fast when you want more people to wait it out. Pro when the garment's hardware and print are the thing you cannot afford to smear, and you have warned them about the wait. You can set a daily generation cap per shopper, tracked by IP, so a broken retry loop does not become a bill. The cap is an abuse control. It is not a speed feature, and a shopper who hits it needs a sentence, not a silent button.

Shoppers who finish a preview are not a random sample of product-page visitors. A wait nobody sits through does not produce a conversion cohort. If your completion rate is poor, you have taps, not a try-on test.

[Shopify virtual try-on and conversion](https://antla.io/blog/shopify-virtual-try-on-conversion) is the commercial frame. Latency is the part of that frame that happens before any conversion can be credited. Put the budget next to the conversion target so nobody celebrates a model setting that shoppers abandon.

## Design progress and retry

Nielsen's point on long operations is practical: show that it has not crashed, show roughly how long remains, and let the person do something else or stop. A percent-done bar is right when you know the work. Try-on generation often does not know a true percent. A fake bar that races to 90% and sticks is worse than words, because it teaches people the bar is a lie. The next time, they leave earlier.

Use a staged status instead:

- "Opening"
- "Uploading your photo"
- "Making the preview. This usually takes about 15 seconds." (Use the number for the path you actually enabled. If you enabled Pro, do not say 15.)
- "Still working" after the typical time, with a cancel
- "Preview ready" or "That one failed. Try again."

Cancel has to cancel. If the job keeps running after they hit back, you will still pay for it, and a late image that pops over a different product is a trust hit. Test the back gesture on iOS Safari. It is the retry button your shoppers already know.

Retry should be one control, and it should say whether you are reusing their photo. Making them hunt the camera roll again, after a failure that was your timeout, is how you turn one error into an exit.

Keep the product page alive behind the wait if you can do it without losing the modal. A shopper who can still read the price, the measurements, and the return window is a shopper with something to do during the 10-second edge. A modal that dims the entire page and offers a spinner is a holding pen.

Do not auto-advance them to the cart when the image lands. They need a second to look. Auto-advance feels fast in a demo and feels like a shove on a phone.

Progress copy is also a disclosure. "Making a preview" is accurate. "Calculating your size" is not, because the model is not measuring them. The latency design and the honesty design are the same sentence.

## Measure abandonment

Abandonment here means they started and did not get a painted preview, or they got one and left without the next action you care about. Those are different leaks. Mix them and you will "fix" a slow model when the upload button was the problem.

Instrument four rates, by device, for the same products:

| Rate | Numerator | Denominator | What a drop means |
|---|---|---|---|
| Open rate | Modal opens | Product views | They never found it, or the button was slow to appear |
| Upload rate | Upload success | Modal opens | Terms, permission prompt, or the picker |
| Completion rate | Preview painted | Upload success | Generation time, errors, timeouts |
| Continuation | Add to cart after a painted preview | Painted previews | The picture, the price, or the size question |

Completion rate is the latency metric. Continuation is the merchandising metric. [Product page engagement and conversion quality](https://antla.io/blog/product-page-engagement-conversion-quality) will tell you if the people who stay are any good. A fast preview that creates junk carts is not a win. A slow preview that nobody sees is not a test of image quality.

Split mobile and desktop. A desktop completion rate can hide a mobile timeout for weeks. Split new and returning if the cookie reuse changes the upload step. Otherwise returning shoppers make the upload rate look healthier than a first visit.

Read errors separately from abandons. An `image_generated_error` is a failure you own. A modal close at 8 seconds may be a person who changed their mind, or a person who thought it had crashed. The status text is how you tell those apart next time. If "Still working" reduces closes in the 8 to 20 second band, the wait was legible and too quiet before.

Give the window enough completions. A hero SKU with 40 mobile tries in a week will swing. Use [conversion metrics](https://antla.io/blog/conversion-engine-metrics-shopify-fashion) for the downstream numbers, and do not declare a latency winner from a single afternoon of paid traffic.

Hold the model path constant during the test. If you switch from Fast to Pro in the middle of the week, you did not measure either budget.

## Limits

Lab Wi-Fi is not the subway. Run one pass on a throttled connection before you trust the office numbers. Chrome's device toolbar is a hint, not the phone in the drawer.

Cached repeat views will look fast and must not set the budget for first uploads.

A percent-done bar you cannot compute is a fiction. Prefer staged words over a bar that stalls at 90%.

None of this measures whether the garment fits. A fast wrong image is still wrong. Speed does not promote a preview into a size decision.

## This week's budget

Write the five ceilings on one card. Time one real upload on a mid-range phone, on cellular if you can. If the painted preview lands past the generation number you put in the status line, change the line or change the path. Then read completion rate for seven days before you touch the button color. The color was never the wait.

## For agents

- Markdown: send `Accept: text/markdown` to this URL (and any other HTML page).
- OpenAPI: https://antla.io/openapi.json
- llms.txt: https://antla.io/llms.txt
- Sitemap: https://antla.io/sitemap-index.xml
- Docs: https://antla.io/docs
- CLI: npx antla info (npm package antla)
- Scope: antla.io is an informational marketing and docs site. The Shopify try-on backend is not on this origin. Install the app from https://apps.shopify.com/antla.
