Notes

Generative AI try-on vs an AR filter: what is the difference?

Last updated: 22 August 2026

An AR filter paints colour onto live video in real time. Generative try-on repaints a still photograph, and each one answers a different question.

An AR filter finds your face in the camera feed and paints colour onto the video as it plays, thirty times a second or faster, on the phone itself. Generative try-on sends one still photograph to a model that repaints the image and hands back a new one. Almost everything else that separates the two follows from that single difference.

How each one works

An AR filter starts by tracking your face. The software fits a mesh to it and follows that mesh from frame to frame. Apple's ARKit, for example, describes the tracked face as an ARFaceGeometry, a 3D mesh whose topology stays constant from one instance to the next, and Apple's own page names applying virtual makeup as a thing you would overlay on it. The software then blends colour onto the regions of that mesh and composites it back into the video. At thirty frames a second it has around thirty-three milliseconds to do all of that, which is why the work has to be cheap: a colour blend over the pixels that are already there, with the edges feathered so it does not look cut out. The try-on you meet on a brand's website runs the same idea in a browser with that company's own tracker.

Generative try-on does not follow you frame by frame. It takes one photograph, sends it to a model running on a server, and the model produces a new image of the same photograph with the makeup present in it. The model overlays nothing. It regenerates the pixels, so the makeup is part of the image rather than a layer sitting on top of it.

The two side by side

Looksmith is the generative example in the table below, and it is coming soon to iOS rather than available today.

The consequences you can check

Why does the same lipstick look different every time in an AR filter?

Because the video underneath it does. Your phone runs automatic white balance and automatic exposure, deciding many times a second what counts as white in the room and pulling the whole frame towards that decision. Sit under a warm bulb and the camera drags the warmth out. Stand by a window and it does the opposite. The filter paints the same colour value in both cases, and you see two different lipsticks.

You can test this in two minutes with any branded try-on page. Open the same shade under a tungsten lamp, then walk to a window with it still running. Nothing about the product changed. The number the filter is painting did not change either.

Which one shows the truer colour?

Neither has earned that claim. Treat any try-on that advertises a colour-accuracy percentage as advertising until it publishes the method behind the number. A still photograph does have one white balance instead of a drifting one, so a generated look at least holds still while you study it. But your camera captured that photograph under one set of conditions and ran it through the same pipeline, and what comes back is a plausible picture rather than a measurement of a pigment. Looksmith does not publish an accuracy percentage for colour, because Looksmith has not measured one.

If the exact shade is the decision, the answer is a physical tester on your own jaw in daylight. Both digital methods only narrow the field before you get there.

Why does a brand's AR filter only ever show that brand?

Because the filter was built around that brand's catalogue. Most branded try-on you have used is one of two engines wearing someone else's name. L'Oréal acquired ModiFace in March 2018, and Perfect Corp, which makes YouCam Makeup, became a publicly traded company on the New York Stock Exchange in 2022. The brand supplies the shade list, the engine supplies the tracking, and the shade list is the point of the exercise.

A single brand's filter cannot show you a look built from three different brands, and it will never conclude that the right answer for you is a product it does not sell. Retailer try-on such as Sephora Virtual Artist or Ulta GlamLab spans many brands, but only the ones that retailer stocks. Who actually builds these engines is taken apart in what is virtual makeup try-on.

How long does a generative try-on take?

About a minute in Looksmith. A model is generating a full image rather than tinting an existing one, and that runs on a server rather than in your hand. Nothing about that is interactive. You are trading the live preview for a finished picture, and if the live preview is what you needed, generative try-on is the wrong tool.

Can I save what the try-on shows me?

This is the difference people notice first. An AR filter is a live overlay, so when you close it there is nothing left unless you thought to screenshot it. Generative try-on produces a file. Looksmith returns a 9:16 poster, which is your own photograph with the makeup painted into it, and optionally a silent four-second looping clip, with the products it used listed beside it. Four of those sit side by side on a screen. One of them can go to a friend with a question attached, or wait on your phone until tomorrow, when you will know what you think.

Does either one change my face?

An AR makeup filter is only painting colour, so in principle your face is untouched. In practice the apps that host those filters often ship reshaping tools alongside them, and a setting left on in one session is easy to miss in the next. Read the app's own feature list rather than assuming the filter is only doing colour.

Generative try-on has the opposite risk. The model is regenerating the whole image, so it can drift unless its builder constrains it. By design, Looksmith will not change face geometry of any kind. A monolid stays a monolid. Your age, pose, expression and framing stay as they were, because what comes back is your own photograph retouched rather than a new portrait. Looksmith preserves glasses exactly and never adds them, and it applies no beauty-filter smoothing. The full list of locked features is in will an AI try-on change my face.

Which one to use for which job

Here is the part most comparisons refuse to write.

The short rule: use an AR filter when the question is about shape and placement and you need to move. Use generative try-on when the question is whether a complete look suits you and you want something to look at again tomorrow.

Where Looksmith sits

Looksmith is a generative try-on app from Mirable Labs, Inc., coming soon to iOS. In Looksmith you save a Likeness once. It is a set of selfies that becomes your reusable face. Then you pick a look one of three ways: Style Me puts you in several looks at once, Create Look takes a description in words, and Mirror a Look transfers the makeup from an inspiration photograph. Looksmith builds the look across twelve regions, from lips and liner through eyeshadow, blush and brows, and it comes back as a poster of your own face with the products listed beside it.

The first six looks in Looksmith are free, granted at onboarding. After that Looksmith uses credits, bought in packs, and there is no subscription. The privacy policy says what happens to the photographs you send.

What neither of them settles

Neither tool touches texture, wear or smell. An AR filter cannot tell you whether a formula sits in your fine lines by lunchtime, and a generated look cannot either, because both work from a single moment and the product is not in the room. A generated look is also not a promise that one specific tube produces exactly that pixel.

Both tools do the same job, which is to shorten the list of things worth trying. They shorten it from different ends. The filter rules shades out while you move, and the generated picture tells you whether the assembled thing is yours. Picking the wrong one is how people end up deciding that virtual try-on does not work. The physical steps that come after either of them are in how to try on makeup without buying it.