Best AI image models for photorealism
Which AI image model looks most like a photograph — skin, light and lens behaviour, judged blind on the same prompts.
0 votes · ratings refit nightly · models are ranked after 40 votes in a category · methodology
| Model | |||||||
|---|---|---|---|---|---|---|---|
| collecting |  Muse Image 1.0 Meta · Try it | — | 0 | — | 12.0s | 12 cr $0.012 | 0.0% |
| collecting | — | 0 | — | 14.1s | 50 cr $0.050 | 11.1% | |
| collecting |  Qwen Image 3.0 Alibaba · Try it | — | 0 | — | 12.9s | 36 cr $0.036 | 0.0% |
| collecting |  Qwen Image 3.0 Pro Alibaba · Try it | — | 0 | — | 9.0s | 41 cr $0.041 | 0.0% |
| collecting |  Seedream 5.0 Pro ByteDance · Try it | — | 0 | — | 30.8s | 49 cr $0.049 | 0.0% |
| collecting |  Nano Banana 2 Lite Google · Try it | — | 0 | — | 15.6s | 36 cr $0.036 | 0.0% |
| collecting |  Ideogram v4 Ideogram · Try it | — | 0 | — | 17.4s | 40 cr $0.040 | 0.0% |
| collecting | — | 0 | — | 27.1s | 35 cr $0.035 | 0.0% | |
| collecting |  Kling v3 Omni Kuaishou · Try it | — | 0 | — | 27.0s | 35 cr $0.035 | 0.0% |
| collecting |  Krea 2 Large Krea · Try it | — | 0 | — | 40.1s | 80 cr $0.080 | 0.0% |
| collecting |  Krea 2 Medium Krea · Try it | — | 0 | — | 25.5s | 40 cr $0.040 | 11.1% |
| collecting | — | 0 | — | 4.9s | 80 cr $0.080 | 11.1% | |
| collecting |  ImagineArt 2 ImagineArt · Try it | — | 0 | — | 27.6s | 60 cr $0.060 | 0.0% |
| collecting |  GPT Image 2 OpenAI · Try it | — | 0 | — | 59.2s | 60 cr $0.060 | 0.0% |
| collecting |  ImagineArt 1.5 Pro ImagineArt · Try it | — | 0 | — | 51.5s | 55 cr $0.055 | 0.0% |
| collecting |  Recraft v4 Pro Recraft · Try it | — | 0 | — | 17.6s | 300 cr $0.30 | 12.5% |
| collecting |  Recraft v4 Recraft · Try it | — | 0 | — | 12.9s | 45 cr $0.045 | 0.0% |
| collecting |  Midjourney Niji 7 Midjourney · Try it | — | 0 | — | 100.6s | 160 cr $0.16 | 0.0% |
| collecting |  Midjourney v7 Midjourney · Try it | — | 0 | — | 135.4s | 160 cr $0.16 | 0.0% |
| collecting | — | 0 | — | 23.5s | 28 cr $0.028 | 0.0% |
How the rating works
Visitors see two blind images for the same prompt and pick the better one, up to 30 votes a day. Every night the votes are refit as a Bradley-Terry model with a bootstrap interval; the interval column is the 95% range. Ties count half.
Speed, price, failures
Measured from the arena's own generations — the same prompt, default settings and resolution for every model — so the numbers are comparable. Price is the standard credit price per 1:1 image; 1,000 credits are one dollar.
Who competes
The 20 most recently released text-to-image models on BudgetPixel. New launches join automatically; the oldest model retires from voting but keeps its page and record.