OutfitGen Research
The benchmark for AI photo editing models, on the edits people actually make
OutfitGen runs the leading image-editing models in production and re-tests them every few weeks. These are the write-ups: the same photos and instructions sent to every model, judged with the names hidden, with every output published so the call can be checked by anyone.
- Published studies
- 3
- Models compared
- 14
- Edit cases
- 23
- Raw outputs published
- 229
AI Image Editing Leaderboard · snapshot 2026-10
Full leaderboard| # | Model | Edits done right |
|---|---|---|
| 1 | Flux 2 Flash · Black Forest Labs | 17 / 18 |
| 2 | Flux 2 Turbo · Black Forest Labs | 17 / 18 |
| 3 | Seedream 5 Flash · ByteDance | 17 / 18 |
| 4 | Seedream 5 Pro · ByteDance | 17 / 18 |
| 5 | Flux 2 · Black Forest Labs | 16 / 18 |
Latest study
October 19, 2026 · Version 1.0
Which AI model copies a garment from a product photo best? A virtual try-on benchmark
13 image models, including the dedicated try-on model FASHN Try-On v1.6, each put the same 5 garments from product photos onto 9 photos of people, three times per case, judged blind on the garment's colour, pattern, logo and cut and on keeping the person.
Read the study
How every study works
The rules are the same for every study, so results stay comparable from one model generation to the next.
01
Identical inputs, production path
Every model receives the same photo and the same instruction, sent through the request path OutfitGen uses for paying customers. No model gets a tuned prompt.
02
Repeated, because these models are not deterministic
At least three runs per model per case. A single output can flatter or damn a model by luck, and a two-run read has already produced a false result in our own work.
03
Judged blind, then unblinded
Contact sheets are shuffled per case and verdicts are written before the key is opened. The unblinded verdicts, and every image behind them, are published.
04
Faces built to expose drift
Portrait cases use subjects with specific identity markers at non-frontal angles, because ordinary stock faces are exactly what edit models regenerate cleanly.
05
Stated, dated, versioned
Each study carries its design table, its limitations, a version and a changelog, and a stable URL to cite. A refreshed finding is a new version, never a silent edit.
06
Proven on real traffic before it ships
A model that wins a study earns a live trial on a slice of OutfitGen generations. Refusal behaviour on real customer photos only shows there, and it has reversed a study result before.
Earlier studies
September 23, 2026 · Updated October 19, 2026
Over-refusal in AI image editing: a benchmark on ordinary swimwear, activewear and lingerie
The same 18 everyday swimwear, activewear and catalog-lingerie edits, sent to 13 current image-editing models five times each, Nano Banana 2.1 included. How often each model refused, and whether the images it returned did the edit.
September 14, 2026 · Updated September 30, 2026
Which AI image model keeps your face? Nano Banana 2 vs GPT Image 2.5 Flare vs Seedream 5 Pro
Four AI image-editing model setups given the same 11 photo edits, three runs each, judged blind: 132 outputs scored on identity preservation, scene preservation, reference-garment transfer and relative speed.
Why we publish this research
We want OutfitGen to be the best AI photo editing experience we can build, which means testing every new model carefully before our users see it. Much of today's generative image technology was built in the open, so we think it is right to give something back. We publish these studies as original research that anyone can reference, cite, criticize and build on.
Studies are free to cite and reuse with attribution. Each one carries a stable URL, a version and a citation block. Questions, corrections and requests for a model to be included go to the address on the about page.
See the winning models on your own photo
OutfitGen runs them for you. Upload a photo, describe the change, no account needed to start.
Try the AI Clothes Changer