RealBench V1
A visual Turing test for image models: can their output pass as a real photograph? Each model's realism score is the share of human votes that judged its AI images as real — voted on by the ImageBench community in a quick real-vs-AI game.
Realism leaderboard
Sorted most photoreal first. Models need at least 10 votes to appear. Today's most convincing model is fal/google/nano-banana-pro.
| # | Model | Realism score | Rated real / votes | Images |
|---|---|---|---|---|
| 1 | fal/google/nano-banana-pro | 41% | 442 / 1,081 | 85 |
| 2 | local/z-image-turbo-6b | 35% | 274 / 779 | 61 |
| 3 | fal/google/nano-banana-2 | 35% | 381 / 1,085 | 85 |
| 4 | bfl/flux-2-max | 33% | 795 / 2,388 | 189 |
| 5 | bfl/flux-2-pro | 27% | 287 / 1,069 | 85 |
| 6 | fal/bfl/flux-1-dev | 25% | 204 / 828 | 65 |
| 7 | fal/fast-sdxl | 25% | 16 / 65 | 5 |
| 8 | openai/gpt-image-2 | 23% | 248 / 1,086 | 85 |
| 9 | bfl/flux-2-flex | 20% | 168 / 822 | 65 |
| 10 | local/sana-1.5-1.6b | 7% | 56 / 825 | 65 |
10,028 votes · snapshot from 6/15/2026