AT A GLANCE
This review covers GPT Image 2’s measured reference fidelity, text and layout handling, and limitations with angle instructions. Compare those results with the requirements of your image.
On this page
Of the three image models we measured, GPT Image 2 was the only one that held a person's features from a reference. It was also the cheapest and emitted the most images per run. Below is only what we confirmed by running it.
Measured
| Item | Measured value |
|---|---|
| Latency | ~24s (22.1 / 25.0 / 25.7s) |
| Cost per image | $0.0174–0.0229 |
| Output | PNG |
| Measured on | 2026-08-28 (via OpenRouter) |
From a single reference it preserved even mole placement across changes of angle, expression, and outfit. For putting a named person into a new scene, it is currently the first choice.
Accepted parameters
| Parameter | Accepted values |
|---|---|
| Aspect ratio | auto / 1:1 / 3:2 / 2:3 / 4:3 / 3:4 / 16:9 / 9:16 / 21:9 |
| Quality | auto / low / medium / high |
| Images per run | 1–10 |
| Reference images | up to 16 |
| Resolution | Not supported — the model has no resolution parameter |
What high quality costs
At quality=high and 16:9 we measured $0.1299 and 72.3 seconds — roughly 6.8× the cost and 3× the latency of auto. There is little reason to use high outside the final image.
Good for
- Putting the same person into a new scene
- Multi-panel layouts and captioned instructions
- Generating several candidates and choosing (up to ten per run)
There is no resolution parameter here. Porting code from another model tends to send one anyway.
Bad for
- Precise camera angles — 30° and 60° drift back toward frontal
- Swimwear, underwear, and similar subjects the provider's safety system rejects
If swimwear is refused, does rewording get through?
It does not. The call fails before generating. Pick a different model for those subjects.
When you need angles
Nano Banana 2 follows angle instructions far more literally — 30° and 60° came out at those angles. It breaks down on multi-panel layouts and embedded text, so a single project may need both models at different stages.
Frequently asked questions
- How many images can GPT Image 2 return per run?
- One to ten — the highest of the three models we measured, which makes candidate-and-choose workflows practical.
- Can I set the resolution?
- No. The model has no resolution parameter at all. Use aspect ratio and quality instead.
- Why was my generation refused?
- Swimwear, underwear, and similar subjects are rejected by the provider's safety system before generation. Rewording does not get through; you need a different model.
