Image Benchmarks: See the Capabilities of Every Model
By Steven Van ·
We ran 39 image models through 15 deliberately hard prompts and put every result on one page.
OpenRouter has published Visual Image Benchmarks, a page comparing all 39 image models offered on OpenRouter against 15 prompts chosen to be hard for image generators. Every model's output sits in a shared grid, sortable by price and generation time, so results can be compared side by side rather than judged from cherry-picked samples.
The prompts are grouped into seven families:
- Improbable scenes, such as a wine glass filled level with the rim
- Counting, such as an exact number of fingers, cards or dice
- Text, including one long exact string on a poster and multiple languages in the same frame
- Spatial relations, covering occlusion and mirror reflections
- Negation, such as a zebra with no stripes
- Editing, testing minimal-diff changes like object or person removal from a reference image
- Consistency, testing whether a model can hold a product or subject steady across a new scene
OpenRouter says it plans to keep the benchmarks updated as new image models are added and to extend the approach to other modalities. Models can be tried on other prompts through OpenRouter's image generation API or Chat.
OpenRouter
One API for 500+ AI models across 80+ providers — pay with credits that work anywhere, with automatic fallback when a provider goes down.
View OpenRouter →