ChatGPT Virtual Try-On: What It Can and Can't Do

ChatGPT can now show you wearing a jacket before you buy it. OpenAI rolled virtual try-on out globally on October 1, 2026, inside ChatGPT's shopping results, running on the Images 2.5 model. You upload one selfie, tap Try On, and get a picture of yourself in the item.

It works better than you expect for colour and silhouette. It is close to useless for fit, and it is not a photo of you in any meaningful sense. Both of those are worth knowing before you screenshot the result and put it anywhere.

What ChatGPT Virtual Try-On Actually Does

ChatGPT virtual try-on is a shopping feature that composites a product onto a photo you upload, so you can preview how a garment or accessory reads against your own skin tone, hair and build.

Three ways to trigger it:

1. Find clothing in ChatGPT's shopping results and tap the new Try On button on the listing.
2. Upload a screenshot of any product from any site and ask ChatGPT to try it on you.
3. Point it at an outfit someone else is wearing in a photo, and it will identify buyable pieces and suggest matches.

The companion launch is Favorites, a library where saved products sit next to the try-on images you generated for them. That part is the genuinely useful bit. Comparing four coats on your own body across a week of browsing is a thing no retailer's own try-on widget lets you do.

The timing is not an accident. Apparel returns run 23 to 25% in the US, roughly three times the in-store rate, and fit and sizing drive about half of them. Sixty-three percent of shoppers now order multiple sizes on purpose. Anything that shaves a point off that number is worth billions to retail, which is why Google shipped a comparable feature a year earlier.

The Disclaimer OpenAI Wrote Itself

Read this before you trust a preview:

Try-on images may not represent the product or your appearance exactly and do not guarantee fit or size.

That is OpenAI's own wording, and it is doing a lot of work. The model has no measurements for you and no size chart for the garment. It is producing a plausible picture, not a simulation. A dress that renders beautifully on your generated body can still arrive two sizes wrong.

What the preview is genuinely good for:

• Colour against your skin tone. Does that olive green work on you, or does it drain you out? The model gets this close to right.
• Silhouette at a glance. Cropped versus longline, wide leg versus tapered.
• Pattern scale. Whether a check is loud or subtle at your height.

What it cannot tell you: shoulder fit, rise, sleeve length, how a fabric drapes when you move, or whether the waistband sits where you want it. Those are the things that cause the return.

Why Your Face Drifts Between Try-Ons

Generate four try-ons from the same reference photo and look only at the faces. They will not match each other.

One photo gives the model a single angle, a single light direction and a single expression. Turn the body, change the garment, re-light the scene for a studio background, and every pixel the original frame did not contain has to be invented. The jaw softens. The hairline moves. Eye spacing shifts a millimetre, which is exactly the amount a human face notices and cannot name.

This is the same failure that makes AI headshots look different every time, and it is structural rather than a bug in Images 2.5. Reference-image conditioning carries style well and identity poorly. We went through this in detail when comparing face consistency across single-reference models, and the result holds here: one input photo is not enough information to pin a likeness.

Neither column is a fit tool. The difference is what the picture is for.

ChatGPT Try-On Privacy: Two Settings Worth Checking

Your reference photo does not vanish after the generation. It is stored on your account so later try-ons can reuse it, which is convenient and also means a full-body photo of you now lives in a place you were not thinking about when you uploaded it.

Two things to do:

1. Open Settings, Personalization, Reference photos to see, replace or delete the stored image.
2. On a personal account, check Improve the model for everyone. While it is on, uploads can be used to train OpenAI models. Coverage of the launch confirms this is the default behaviour on personal plans.

Nobody needs to be alarmed about this, and nobody should discover it by accident either. If you would not post the photo publicly, think twice before making it your permanent stored reference.

How to Get a Usable Try-On Image

Four things change the output more than the prompt does:

1. Shoot full body, standing, arms visible. A head-and-shoulders selfie forces the model to invent your entire torso, and it will get your proportions wrong in a flattering direction.
2. Stand against a plain wall in daylight. Window light, no flash, no ceiling spots. Patterned backgrounds bleed into the generated garment edges.
3. Wear something fitted and plain. A bulky hoodie hides the shape the model is supposed to dress.
4. Generate the same item twice. If the two previews disagree about the garment, trust neither. If they agree, the colour read is probably reliable.

The one habit worth building: judge colour and silhouette from the preview, then check the brand's actual size chart against your own measurements. The AI handles the part you cannot picture. The tape measure handles the part it refuses to promise.

Where This Leaves Your Profile Photos

Try-on is a shopping tool that happens to produce a picture of you. That is a different job from a photo meant to represent you on LinkedIn, Hinge or a company team page, and the output shows it: product-first crop, flat catalogue lighting, a likeness nobody guaranteed.

The outfit-photo use case is real, though, and it is the reason people keep screenshotting these. If what you want is a set of photos of yourself in different outfits and settings, with your own face held steady across all of them, that comes from a model trained on your photos rather than conditioned on one. Our virtual try-on page does the garment version of this, and a full photo session does the wider one.

Worth knowing if these are headed for a work profile: recruiters spot AI photos far less often than they think, but two thirds say they would be put off once they knew. A photo that looks like you solves both halves of that. A try-on screenshot solves neither.

And if you came here because the ChatGPT headshot route disappointed you, the Images 2.5 headshot problem has the same root cause as the drifting faces above. One reference photo, one invented likeness. Reporting on the launch is right that this is a strong shopping feature. It was never built to be a camera pointed at you.