How Does AI Clothing Try-On Work? The Technology Explained Step by Step (2026)
How AI clothing try-on works: two photos become a still on the Shopify PDP. Not AR, not a size tool. Research papers and a 7 Oct 2026 app table.
AI clothing try-on generates a still of a shopper wearing a garment from two photos: theirs, and the product’s. It is not a live AR overlay and not a size picker. Antla runs that photo flow on a Shopify product page from one upload, with no shopper app and no 3D scan.
Most clothing teams still buy the wrong object. They demo a glasses camera filter, then expect a midi skirt to drape. They buy a size quiz, then expect a neckline to resolve. AI clothing try-on is a third job: a generated still of this person in this SKU, delivered where the add-to-cart decision happens.
Photo try-on is a still of the shopper in the product, not a live camera sticker and not a tape measure.
The label “virtual try-on” hides three different products
Shopify’s 2026 virtual fitting room guide stacks AR overlays, AI, and 3D modeling under one retail phrase. That is useful as a map. It is a bad buying spec. A live face overlay, a generative clothing still, and a rotatable 3D mannequin answer different shopper questions and need different catalog work.
Live AR tracks a face, hand, or body in the camera and composites a 3D or 2D asset onto the feed. It is fast for glasses, makeup, and some jewelry because those objects sit on a stable landmark. It is a weak dress preview. A skirt does not become fabric because you tracked a torso box.
3D avatar rooms dress a generic or measured body in simulated garments. Shopify’s same guide treats 3D visualization as its own layer: you need garment meshes, not just PDP JPEGs. That pipeline is why mid-size catalogs stall. You are not installing a widget. You are standing up an asset factory.
Generative photo try-on takes a shopper image and a garment image and synthesizes a new still. Google’s 2023 write-up of shopping try-on is explicit about the break from older “cut, warp, paste” methods: those leave misplaced folds and garments that never quite sit on the body. Diffusion generates the clothing pixels instead of sliding a cutout around.
Google’s public feature puts garments on models inside Shopping results. A Shopify clothing app puts the still on your product page, on the shopper who is about to buy. Same family of image math. Different place in the funnel. The chat-app version of that split is in ChatGPT virtual try-on vs a store PDP.
If you sell apparel on Shopify, you want the photo still. You keep a size chart or size app for the label. You do not owe anyone a 3D closet.
Two photos in, one still out
I would walk a founder through the job in this order, because this is the order the model actually sees.
1. The shopper image is the body the pixels have to respect
The shopper taps Try on, then uploads a photo already on the phone or takes one in the browser. There is no separate consumer app to download. A clear, front-facing frame helps: standing, arms slightly away from the torso, garment area visible, ordinary indoor light. A cropped selfie can be enough for a top. A maxi dress needs more of the body in frame.
The system has to find a person in that photo. Research pipelines segment the figure, estimate pose, and build a “clothing-agnostic” view of the body so the new garment is not fighting the old one. You do not need to explain U-Nets to a merchandiser. You do need to know this: a hunched side angle, a packed subway selfie, or a toddler covering the waist will give the model less to work with.
Photo try-on is not a body scan. It does not output bust, waist, and hip in centimeters. It reads a picture.
2. The garment image is your existing product photo
You do not 3D-scan the dress. You do not upload a DXF. Photo try-on reads the catalog image the PDP already has: usually a front packshot or a front on-model shot where the construction is visible. Neckline, sleeve, print, closure, hem. A back-only hero, a cropped texture detail, or a lifestyle crop that hides the silhouette is a bad control image.
That is why try-on and photography are partners, not substitutes. Studio images still answer “what is the product?” The generated still answers “how might this look on me?” The split is in product photography vs AI virtual try-on.
If the first gallery image is the back of a bomber, pick a front image in the app. Do not hope the model invents a chest graphic that is not in the photo.
3. Generation writes a new picture. It does not paste a sticker
Older methods warped the garment onto a silhouette, then blended. TryOnDiffusion (Zhu et al., 2023) describes the newer pattern in public: two images in, a diffusion process that warps and blends in one pass, garment detail preserved while pose and shape change. Google’s shopping post says the same thing in plainer language: each image goes through its own network, they share information through cross-attention, and the output is a photorealistic still of the person wearing the garment.
Two 2024 papers in the same family, IDM-VTON and OOTDiffusion, push garment identity (prints, logos, construction) so the result looks like that SKU, not a cousin. Commercial Shopify apps are not those research checkpoints. Do not treat an arXiv figure as a product spec. The useful takeaway for a merchant is simpler: the model is generating pixels from two photos. It is not running a physics simulator of cotton vs silk, and it is not measuring the shopper for a size.
Antla does not publish a median generation time. Anyone quoting “10 to 30 seconds” as an Antla number made it up. The shopper waits on a still. They are not watching a live overlay.
4. Shopify delivery is a product-page widget, not a new storefront
The still has to land where hesitation lives: on the product page, next to the gallery and add to cart, on mobile first. That is a theme app block, not a Liquid rewrite. The feature surface is Antla virtual try-on.
What merchants see after launch is usage, not a fitting-room video. Which SKUs get tried on, who completes a preview, who then buys. Across more than 500,000 Antla try-ons, users who complete a preview convert at 3.8%, a try-on-user rate, not store-wide conversion. Source and the denominator warning: Shopify virtual try-on conversion.
Photo try-on vs AR vs 3D, checked 7 October 2026
Listings change. These cells are from the Shopify App Store on 7 October 2026.
| Job on the store | Shopper input | Catalog input | Shopify example | Entry paid plan (7 Oct 2026) |
|---|---|---|---|---|
| Clothing photo try-on | One photo, in the browser | Front product photos | Antla (Built for Shopify) | $19.99/month (Trend: 100 try-ons, then $0.16). 7-day trial, no free plan |
| Glasses live try-on | Device camera, face tracking | 3D frames | Fittingbox Virtual Try-On | $59/month (Bronze: up to 10 active products). 14-day trial |
| Jewelry / accessories AR | Live camera on face, hand, or body | Images or optional GLB | Camweara Virtual Try On | $39/month (Lite: 70 products). No free trial on the listing |
| Makeup, glasses, hair color AR | Live camera | AR cosmetics / frames / dye | Banuba AI Virtual Try-On (Built for Shopify) | Free on development stores only. Paid Basic starts at $319/month (10,000 try-ons). 14-day trial on paid |
I would never evaluate a dress PDP by installing Banuba or Fittingbox. Those apps are honest about face-tracked categories. They are not clothing drape tools. Camweara’s listing names clothes among supported types, and still describes an AR overlay, not a generative garment still. If the demo is a live camera sticker on a t-shirt, you are not looking at the same product as a photo try-on of a wrap dress.
3D avatar rooms rarely show up as a cheap Shopify app for a 200-SKU seasonal drop, because the cost sits in meshes, not in the monthly plan. If a vendor asks you to model every hem, you are back in generation two of virtual try-on technology in fashion.
What the still cannot prove
A good preview can show length against the shopper’s leg, whether a square neck sits wide, how a print hits the chest, how sheer a white blouse reads on their skin. It can also look convincing in a size that still pinches at the waist.
Try-on does not pick a size. Antla has no size chart or size recommendation of its own. It sits beside whatever Shopify size app you already run. The listing line is accurate: size charts tell customers what fits; Antla shows them how it looks. Mixing those jobs is how stores file “looked different on me” as a sizing ticket.
Sheer fabric, extreme drape, heavy layering, and odd poses still fail more often than a straight-on cotton tee. Does virtual try-on work? is the evidence page. This page stays on mechanism.
The commercial reason to care is not a cinematic render. It is the expectation gap that sends apparel back. Shopify’s 2026 returns guide cites the NRF and Happy Returns 2025 landscape: an estimated 19.3% of online sales were returned that year, against 15.8% for retail overall. Apparel and footwear sit above the online average because fit is hard to judge on a screen. A preview does not delete that rate by itself. It answers a visual question the size chart never could.
Baymard’s apparel UX research still treats fit communication, product visuals, and returns UX as separate problems. Try-on does not replace measurements, model stats, or honest photos. It fills the personal visualization hole those assets leave.
What happens to the shopper’s photo
Quote the live policy, then stop. Per Antla’s privacy policy:
Images are deleted and not accessible within 72 hours.
Do not tell shoppers the photo is never stored. Do not claim it is unused for training unless the policy says so. It does not. Put that 72-hour line in your own store policy if you run photo try-on, and keep the terms toggle on if you want consent before generation.
Antla is Built for Shopify certified. It is not SOC 2 or ISO certified.
A clothing merchant’s working checklist
- Name the job. If the return file is too-small / too-big, fix the chart or size app first. If it is “not as pictured,” you want a photo still.
- Start on one family. I would start with a structured top or a dress that has a readable neckline, not a sheer overlay or a piled-on lookbook set.
- Give the model a front garment photo. Swap the control image when
images[0]is a back shot. - Put Try on next to the gallery, on mobile, not in a tab under the description.
- Leave Find my size next to the selector. Two buttons. Two questions.
- Run the flow on your own catalog before you argue about theme code. The public demo is linked at the end of this page.
- Measure try-on users separately from store-wide conversion. Same rule as the 3.8% cohort: do not mix denominators.
Antla plans on 7 October 2026, from antla.io/pricing: Trend $19.99/month (100 try-ons, then $0.16), Runway $49.99/month (500, then $0.12), Unlimited Fashion $199.99/month (2,000, then $0.09). Every plan includes a 7-day free trial. Klaviyo is on the published plans. There is no free plan.
Questions merchants ask after the demo
How does AI clothing try-on work?
It takes a shopper photo and a product photo, then a generative model writes a new still of that person wearing that garment. The still is placed on the product page. It is not a live camera overlay, a 3D avatar, or a size recommendation.
Do I need 3D models of my garments?
No. Photo try-on uses the product images you already shoot for the PDP. 3D meshes belong to avatar and some AR pipelines, not to this flow.
How is AI try-on different from AR try-on?
AR composites a digital object onto a live camera feed, which is why it is common for glasses and makeup. AI clothing try-on generates a still from two photos, which is why it can show a dress on a body without a 3D asset of that dress.
How long does a try-on take?
Antla does not publish a median generation time. The shopper waits for a still in the widget. Treat vendor “instant” copy as marketing until you time it on your own catalog.
Does it work from one shopper photo?
Yes. The Shopify listing describes try-on from a single photo. A fuller body in frame helps longer garments. Shoppers use a photo already on the phone. They do not install a consumer app.
Does AI try-on show whether clothes will fit?
No. It previews look: length, silhouette, coverage, color on the shopper. It does not measure the body or guarantee a size. Keep a size chart or size-recommendation app next to the selector.
What happens to the customer’s photo?
Antla’s privacy policy states: “Images are deleted and not accessible within 72 hours.”
Does Antla recommend a size?
No. Antla has no size chart or size recommendation of its own. It works alongside any Shopify size app.
Is Antla SOC 2 or ISO certified?
No. Antla is Built for Shopify certified and holds neither SOC 2 nor ISO certification.
The adjacent explainer on this site, what virtual try-on is for fashion stores, stays live. This URL is the clothing-technology page: two photos, a still, a Shopify PDP.
Try it on a real catalog: Antla.io/demo
If you are past the mechanism: How to add virtual try-on to Shopify · AI virtual try-on in ecommerce
About the author: Aaron founded Antla. Clothing try-on is two photos and a still on the product page. If the demo needs a live camera and a pair of glasses, you are not looking at a dress tool.