Which AI model copies a garment from a product photo best? A virtual try-on benchmark

Why this study

Research question

Given a photo of a person and a product photo of a garment, which image models put that exact garment on that person, with its colour, pattern, logo and cut intact, without changing who the person is or how they stand?

Models

Rows are in rank order: try-ons done right, then median time per image. GPT Image 2.5 Flare ran at two quality settings, and each setting counts as a model. Seedream 5 Lite, on the leaderboard, was not part of this study. OutfitGen Engine is OutfitGen's own system, shown last and not ranked. Every single model was called through fal's hosted endpoints on 19 October 2026.

Scorecard

A try-on counts as done right when every image the model returned copied the garment (colour, pattern, logo or text, and cut) onto the same person in the same pose. ✓ means every case or run; ! means some; ✗ means a large miss. Times are relative to the fastest model in this session, which ran slower than the leaderboard's September session, so compare them within this table only. Ties in the ranking are broken by median time, which these rounded ratios can hide.

Charts

Method

1. Five garments were generated once as product photos with fixed seeds, each built around one hard detail: a wordmark and emblem, stripe spacing, colour blocking with a chenille letter, a dress's length and sleeve length, and a large printed number with a small sleeve logo. The varsity jacket was photographed on a man, to test whether a model copies the person in the product photo along with the garment.
2. Nine subject photos covered five synthetic adults: a man and a woman shown front-on, with arms crossed, with a hand flat on the chest, and with long hair over both shoulders, plus a plus-size woman, a heavyset older man and a petite older woman. Each wore a plain top so the swap is unambiguous.
3. Every model received the subject photo, the garment photo and the instruction "Change the outfit in this photo to match the reference clothing image" through one request path, three times per case. Where a maker documents a wording for two-image edits, the instruction was sent in that wording. FASHN Try-On v1.6 takes no instruction and received only the two photos.
4. Delivered images went into one contact sheet per case with the model columns shuffled. One reviewer judged each column across its three runs without the key: the garment's colour, pattern, logo or text, and cut; whether the person's face, build and skin tone stayed the same; and whether the arms, hand and hair stayed as they were. Verdicts were written before the key was opened.
5. Changes outside the garment, such as trousers swapped for the ones in the product photo, a zoomed-out crop or a letterboxed frame, were recorded but did not fail a try-on.
6. Qwen Image 3 ran with fal's prompt expansion off and a fixed 3:4 size, as in the earlier studies; 7 of its 39 runs timed out at the 90-second limit and are excluded. No other run failed.
7. Grok Imagine 2.0 returned an unrelated outfit in every run, so a follow-up the same day sent it two of the cases with two other wordings, two runs each: one naming the garment, and one in the image-order format other makers document. The follow-up is reported in Findings and is not part of the counts.

Results by case

One representative output per model is shown for each case; the judging used every run, 539 outputs in total.

Logo hoodie, clear view

Run this edit

Instruction: “Change the outfit in this photo to match the reference clothing image”

Look for: Is the ALPINE CO. wordmark spelled right, with the pine emblem, the pocket and full-length sleeves?

Verdict: Ten models and OutfitGen Engine copied the hoodie, wordmark and pine emblem in every run. Flux 2 Turbo and Flux 2 Flash cut the sleeves short in one run each. Grok Imagine 2.0 returned navy shirts and a grey henley instead.

In OutfitGen this is a product-photo job for the AI virtual try-on.

Varsity jacket shown on another person

Run this edit

Instruction: “Change the outfit in this photo to match the reference clothing image”

Look for: Does the jacket arrive on her, without the face, build or trainers of the man in the product photo?

Verdict: The identity trap. Flux 2 Turbo and Flux 2 Flash returned the man from the product photo instead of her in all three runs. Nine models and OutfitGen Engine put the jacket on her with face and pose kept in every run. Nano Banana 2 moved a hand into a pocket in one run, and Grok Imagine 2.0 returned sweaters and a blouse.

In OutfitGen this is a product-photo job for the AI virtual try-on.

Logo hoodie, arms crossed over it

Run this edit

Instruction: “Change the outfit in this photo to match the reference clothing image”

Look for: Do the arms stay crossed with the wordmark showing above them, and do the sleeves reach the wrists?

Verdict: Every model kept the arms crossed in every run. Nine models and OutfitGen Engine also copied the hoodie right every time. Flux 2 Flash drew short sleeves in two runs and Flux 2 Turbo in one, FASHN Try-On v1.6 doubled a sleeve into a dark fold once, and Grok Imagine 2.0 returned a leather jacket, a blazer and a plaid shirt.

In OutfitGen this is a product-photo job for the AI virtual try-on.

Breton stripe top, arms crossed over it

Run this edit

Instruction: “Change the outfit in this photo to match the reference clothing image”

Look for: Do the stripes break around the arms and resume at the right spacing, with long sleeves, a boat neck and the red anchor?

Verdict: Eight models and OutfitGen Engine carried the stripes around the crossed arms in every run. FASHN Try-On v1.6 warped them on the arms and added extra anchors, Qwen Image 3 drew a short-sleeve crew tee, Flux 2 darkened her skin in one run and Flux 2 Turbo added a hem band in one. Grok Imagine 2.0 returned a turtleneck, a blazer and a blouse.

In OutfitGen this is a product-photo job for the AI virtual try-on.

Varsity jacket on a heavyset older man

Run this edit

Instruction: “Change the outfit in this photo to match the reference clothing image”

Look for: Does the jacket fit his build, with his bald head, grey beard and age kept, rather than the slimmer man in the product photo?

Verdict: All three Flux 2 models swapped in the man from the product photo in at least one run: Flux 2 Flash in two of three, Flux 2 and Flux 2 Turbo in one each, Turbo with three garbled R patches. Nine models and OutfitGen Engine fitted the jacket to his own build in every run. Grok Imagine 2.0 returned blazers and a check shirt.

In OutfitGen this is a product-photo job for the AI virtual try-on.

Graphic tee on a heavyset older man

Run this edit

Instruction: “Change the outfit in this photo to match the reference clothing image”

Look for: Does RIVERSIDE 23 read right at his chest's scale, with the small red sleeve logo, and does his pale skin stay pale?

Verdict: Eight models and OutfitGen Engine printed RIVERSIDE 23 and the sleeve logo on him in every run. FASHN Try-On v1.6 changed the typeface, enlarged the sleeve logo and added forearm tattoos in every run. Flux 2 Turbo darkened his skin in all three runs and Flux 2 in one. Grok Imagine 2.0 returned a suit and blazers.

In OutfitGen this is a product-photo job for the AI virtual try-on.

Findings

Reproducing the study

The study is built to be re-run on the same photos whenever a new model ships. The download below has every run with its blind verdict.

• Fixed-seed sources. Every subject and garment photo is generated once with a fixed seed and kept, so a future run compares new models on the same people and the same products.
• Three runs per case per model. A try-on that works once can fail on the next run; a case only counts when every delivered run is right.
• One request path. Every model receives the same two photos and instruction through the same code, so differences come from the models.
• A trap case. The varsity jacket is shown on a different person, which is the quickest way to see whether a model copies only the garment.

Limitations

• The subjects and garments are AI-generated, not shopper photos or real catalog shots. Real product photos vary more in lighting, folds and angle.
• The source photos were generated with Flux 2. That could make them slightly easier for the Flux 2 editors, which still placed in the bottom half.
• Three runs per case bounds luck but not tightly. A model that got a case right 3 of 3 may miss it occasionally at larger volume.
• One reviewer judged every image. The delivery counts do not depend on judgment; the try-on verdicts do.
• Every model got one instruction wording, in its maker's documented format where one exists. Grok Imagine 2.0 was also tried with two other wordings and still did not copy the garment.
• FASHN Try-On v1.6 ran in balanced mode; its slower quality mode was not tested.
• Changes outside the garment, such as trousers swapped for the product photo's, were recorded but not counted against a try-on.
• Times come from one session that ran slower than the leaderboard's September session, so they compare models within this study only.
• Models change without notice. These results describe the models as served on 19 October 2026.

Definitions

Model

One model at one setting. GPT Image 2.5 Flare at medium and at high quality count as two models.

Virtual try-on

Putting a garment from a product photo onto the person in another photo, keeping the person, pose and scene.

Images delivered, of 39 runs

Runs that returned an image, out of 13 cases x 3 runs. No run in this study was refused; the only misses were timeouts.

Clear view, done right of 4

Try-ons done right on the four cases where nothing covers the garment: the hoodie and stripe top on a man, the wrap dress and the varsity jacket on a woman.

Arms, hand or hair over the garment, of 4

Try-ons done right on four cases where crossed arms, a hand flat on the chest or long hair covers part of the garment. The arms, hand or hair have to stay where they were.

Different builds and ages, of 5

Try-ons done right on a plus-size woman, a heavyset older man and a petite older woman, five cases in all. The garment has to fit their body, not the product photo's.

Same person kept, of 13

Cases where every run kept the subject's face, build and skin tone.

Time per image vs the fastest (lower is better)

Median seconds from request to delivered image, divided by the fastest model's median in the same session.

Try-ons done right, of 13 (higher is better)

Cases where every delivered run copied the garment's colour, pattern, logo or text and cut, onto the same person, in the same pose.

Disclosed entry

OutfitGen's own system run on the same inputs and shown for transparency, outside the ranking.

Common questions

How to cite

This study lives at one stable URL. It is updated in place, with a version number and a changelog, so a link made today keeps resolving and a reader can tell which version was read.

@misc{outfitgen-virtual-try-on-2026-10,
  title  = {Which AI model copies a garment from a product photo best? A virtual try-on benchmark},
  author = {{OutfitGen}},
  year   = {2026},
  month  = {October},
  note   = {OutfitGen Research, version 1.0},
  url    = {https://outfitgen.ai/research/virtual-try-on-2026-10}
}

Version history

1. Version 1.0, October 19, 2026. First publication. Thirteen models plus OutfitGen Engine, 13 try-ons, three runs each, 546 requests.

References

• Choi et al., VITON-HD (CVPR 2021). High-resolution virtual try-on with a dedicated network and a public dataset of 1024x768 try-on pairs.
• Morelli et al., Dress Code (ECCV 2022). A multi-category virtual try-on dataset covering upper-body, lower-body and dress garments.
• FASHN v1.6 announcement. FASHN's release post for the try-on model tested here.
• Black Forest Labs. Maker of the Flux 2 models.
• ByteDance Seed. Maker of the Seedream 5 models.
• Qwen (Alibaba). Maker of Qwen Image 3.
• Google DeepMind. Maker of Nano Banana 2 and Nano Banana 2.1.
• OpenAI. Maker of GPT Image 2.5 Flare.
• xAI. Maker of Grok Imagine Image 2.0.
• Meta AI. Maker of Muse Image.
• fal. The inference platform every model in this study was called through.