# AI virtual try-on: tested on a real outfit, plus clothes on a model for your store

> AI virtual try-on takes two photos — a person and a garment — and “dresses” the person while keeping their face and pose. We tried an oversized cardigan on in Nano Banana Pro and GPT Image 2.5, then, for sellers, turned the same cardigan into catalog shots on two different body types and an autumn lifestyle photo. Below: prompts, a model comparison and what try-on still can't show.

Published 2026-10-03 · Updated 2026-10-03
URL: https://ai2face.com/blog/odezhda-na-modeli-nejroset

AI virtual try-on solves two different problems. For shoppers, it shows how a piece will look on them before they order. For sellers, it replaces a photo shoot with a model: a photo of a garment on a hanger becomes a catalog shot and a lifestyle image. We tested both scenarios on one oversized cardigan — a tough case with a chunky cable knit, a loose cut and wooden buttons.

## 1. How AI virtual try-on works

Image generation models accept several references at once. Try-on uses two: a photo of the person and a photo of the garment. The prompt needs to name the role of each — that's what OpenAI advises for GPT Image (“reference each image by index and description”) and what Google recommends for [Nano Banana](/models/nano-banana) (several images as context for a new scene).

The prompt rests on three parts:

1. **Reference roles:** “image 1 is a woman, image 2 is a cardigan”.
2. **What to keep:** face, hair, body proportions, pose, the rest of the outfit, background.
3. **What the garment looks like:** color, texture, cut, hardware — so the model doesn't “simplify” it.

## 2. Test: trying on a cardigan on a full-length photo

The person photo is full-length against a light wall, in a white T-shirt and jeans. The garment is a mustard chunky-knit cardigan on a hanger. We sent the same prompt to [Nano Banana Pro](/models/nano-banana-pro) and [GPT Image 2.5](/models/gpt-image).

```
Image 1 is a woman, image 2 is a cardigan. Dress the woman from image 1
in the cardigan from image 2, worn open over the white T-shirt. Keep her
face, hair, body proportions, pose, jeans, sneakers and the gray wall
exactly the same. Keep the cardigan exactly as in image 2: mustard,
chunky cable knit, oversized, large wooden buttons, patch pockets,
dropped shoulders. Natural fit and folds, photorealistic, no text,
no logos
```

![Four frames: a full-length woman in a T-shirt and jeans, a mustard cardigan on a hanger, the try-on in Nano Banana Pro and the try-on in GPT Image 2.5](https://ai2face.com/_astro/tryon.BCE8LuQU_gLYMY.jpg "Left to right: person photo, garment photo, Nano Banana Pro, GPT Image 2.5. The person and garment were generated for this test — in real use, upload your own photos")

| | Nano Banana Pro | GPT Image 2.5 |
|---|---|---|
| Face, pose, background | preserved | preserved |
| Cardigan length | shorter than the original | closer to the original |
| Oversized fit | became closer-fitting | loose cut preserved |
| Knit and cables | present, finer | present, closer to the original |
| Pockets and buttons | in place | in place |

Our takeaway: for trying on a specific garment, GPT Image 2.5 was more accurate on the cut, while Nano Banana Pro slightly “tailored” the oversized fit. Neither model touched the face or figure, and that's what matters most for trying clothes on your own photo.

## 3. How to prepare photos for try-on

- **Person:** full-length, facing the camera, against a plain background, in even daylight. Arms slightly away from the body so they don't cover the torso. Clothing in the photo should be simple and close-fitting: a T-shirt, jeans.
- **Garment:** whole, from the front, without wrinkles, on a hanger or laid flat on a plain background. The better the cut and details are visible, the more accurate the try-on.
- **Prompt:** list the garment's details in words — color, texture, length, hardware. The model checks the image against them.

If the result is almost there, fix it with one sentence: “make the cardigan longer, to mid-thigh” — the model will keep everything else.

## 4. AI clothing on a model for your store

A seller doesn't need their own full-length photo: the model can be described in words. From the same photo of the cardigan on a hanger, we made catalog shots on two different body types.

```
Marketplace catalog photo: a slim young model with light blonde hair
in a bun stands full-length against a seamless light warm gray
background, wearing this cardigan buttoned over a white top, straight
beige trousers and white sneakers, one hand in a patch pocket. Keep the
cardigan exactly as in the reference: mustard, chunky cable knit,
oversized, large wooden buttons, patch pockets. Soft even studio light,
photorealistic, no text, no logos
```

![The same mustard cardigan on two models in a studio: a slim model with a bun and a plus-size model with curly hair](https://ai2face.com/_astro/studio.DhfkdudY_ZRXVCN.jpg "One garment, two body types — Nano Banana Pro. In the second prompt we changed only the model description. The cardigan, buttons and pockets match")

Showing a piece on different body types helps shoppers understand the cut — that used to take two models and two shoots. For lifestyle shots, change only the setting and the outfit:

```
Lifestyle photo: a young woman with light blonde hair walks through
a park with golden autumn leaves on a sunny morning, wearing this
cardigan open over a cream turtleneck and dark jeans, holding a paper
coffee cup without a logo, a genuine smile. Keep the cardigan exactly
as in the reference. Photorealistic, natural light, no text, no logos
```

![A young woman in a mustard cardigan and cream turtleneck walks through an autumn park with a cup of coffee](https://ai2face.com/_astro/lifestyle.DkTiibB0_1f7ej6.jpg "Lifestyle — Nano Banana Pro. Same garment; the scene took two lines of prompt")

The main image, infographics and the other listing slides are covered in our overview of [AI for ecommerce sellers](/blog/nejroseti-dlya-marketplejsov), and ready-made prompts are in [AI product photography prompts](/blog/promty-dlya-kartochek-tovara).

## 5. The same model across all your listings

If your store needs one recognizable face on every product, describing the model in words each time doesn't work well: their appearance will drift. AI2Face has a builder for that: you create a character once from 1–4 photos, then assemble a shot from the character, clothing, location and pose. The character is saved and dropped into any scene without re-uploading. More on shooting people in our guide to [AI photoshoots](/blog/ii-fotosessiya).

## 6. Virtual try-on in ChatGPT

On October 1, 2026, OpenAI added virtual try-on to ChatGPT: clothing and accessory cards in shopping results got a “Try on” button. The user uploads or takes a selfie, and ChatGPT Images generates an image with the item. The reference photo is saved for future try-ons and can be changed or deleted in personalization settings. It's a feature for shoppers inside ChatGPT; for shooting their own catalog and listings, sellers are better off working with the models directly through prompts.

## 7. Limitations of AI try-on

- **It doesn't show size or fit.** AI draws how a piece looks, not how it will fit your measurements. Size only by the seller's chart.
- **Loose cuts get “tailored”.** Models sometimes make oversized fits, volume sleeves and length more compact — Nano Banana Pro did so in our test.
- **Small details change.** Number of buttons, pocket shape, fabric pattern — check against the real garment.
- **You can't feel the fabric.** Weight, softness and sheerness come across only approximately.
- **Good source photos matter.** Waist-up shots or bulky clothing give noticeably worse results.

## Bottom line

AI virtual try-on already reliably keeps the face, pose and background and renders the garment close to the original. In our test GPT Image 2.5 was more accurate on the oversized cardigan's cut, while Nano Banana Pro is the better fit for catalog shots and lifestyle images with models. For shoppers, try-on helps judge color and style, but not size. For sellers, it replaces a shoot with a model: one garment on a hanger becomes a catalog on different body types and real-life scenes. In AI2Face, Nano Banana Pro and GPT Image 2.5 are in one generator, and the character builder keeps the same model across all your listings.

## FAQ

### Which AI is best for virtual try-on from a photo?

In our test both models kept the face, pose and background. GPT Image 2.5 reproduced the length and knit of the oversized cardigan more accurately; Nano Banana Pro made it shorter and closer-fitting. For shots with people in general, Nano Banana Pro is more convenient: its hands and faces look more natural.

### Can I try on clothes with AI for free?

Partly. AI2Face gives you starter credits when you sign up — enough for a few try-ons. The cost of each generation is shown before you run it.

### What photo do I need for AI try-on?

Full-length, facing the camera, against a plain background, in even light, arms slightly away from the body, in close-fitting clothes. The garment should be shown whole, from the front, on a plain background or a hanger.

### Will AI show whether the size fits?

No. AI shows color, style and how the piece works with an outfit, but not the fit for your size or how the fabric feels. Check size against the seller's size chart.

### Can AI put clothes on a model for a product listing?

Yes. Upload a photo of the garment and describe the model, pose and background — the AI will create a catalog shot. Check buttons, pockets and fabric patterns: the model may change small details.

## Sources

- https://help.openai.com/en/articles/6825453-chatgpt-release-notes
- https://developers.openai.com/api/docs/guides/image-prompting
- https://ai.google.dev/gemini-api/docs/image-generation