How this leaderboard is measured
Every model generates images for the same 10 prompt groups. Outputs are compared across 1,050 head-to-head matches scored by an ensemble of six vision-language judge models (PickScore, HPSv2, ImageReward, VQAScore, a VLM judge and CLIPScore). Models are ranked by mean per-group Elo, cross-checked with a Bradley–Terry estimate.
- What it is: a transparent, reproducible proxy for image quality that we re-run whenever a major new model launches.
- What it isn't: human preference. Scores come from automated judges, so read this alongside human-preference arenas — not as an absolute verdict.
- Tracks: two are live — text-to-image and image editing (image-to-image); switch between them with the toggle at the top. Upscaling and restoration publish as their research completes.
AI & methodology disclosure:rankings are generated by automated vision-language scorers, not human raters. Pixazo hosts these models but does not own them — each model's provider is named in the table. Benchmark designed and reviewed by Deepak Joshi, AI Research, Pixazo. Last updated June 2026.
Frequently asked questions
Which AI image generation model is best in 2026?
In Pixazo's image generation benchmark, gpt-image-2 (medium) by OpenAI leads with a mean Elo of 1324 across 10 real prompt groups and 1,050 pairwise matches. gpt-image-1.5 and Microsoft's MAI-Image 2.5 follow within a few Elo points, so the top of the board is a tight pack rather than a runaway.
Are open-source AI image models competitive in 2026?
Most frontier image generation models are still proprietary. Among the 15 models tested here, NVIDIA Cosmos 3 Super is the open-weight option, while OpenAI, Google and ByteDance hold the top spots. You can filter the ranking table by open vs. closed source.
How is this AI image generation ranking calculated?
Every model generates images for the same 10 prompt groups. Outputs are compared head-to-head across 1,050 pairwise matches, scored by an ensemble of six vision-language judges (PickScore, HPSv2, ImageReward, VQAScore, a VLM judge and CLIPScore). Results feed a per-group Elo rating; the headline score is the mean per-group Elo, cross-checked with a Bradley–Terry estimate.
Why use AI judges instead of human voters?
Automated vision-language scorers make the benchmark fully reproducible and let us re-run it the moment a new model ships, without waiting for thousands of human votes. It is a proxy for human preference, so we publish the full per-judge and per-group breakdown rather than a single number.
What does “Elo per dollar” mean?
It is a model's mean Elo divided by its Pixazo per-image price — a value indicator, not a quality score. A model with slightly lower quality but far lower cost can deliver more Elo per dollar than the outright leader.
Which model offers the best value for the price?
On Pixazo's real per-image rates, grok-imagine-image ($0.020) returns the most Elo per dollar, and NVIDIA Cosmos 3 Super offers strong quality at a low price. The most expensive model costs more than 10× the cheapest.
How often is the leaderboard updated?
The board is re-run whenever a major new image model launches on Pixazo, and the update date is shown at the top of the page. Two tracks are live — text-to-image and image editing (image-to-image), switchable at the top; upscaling and restoration tracks publish as their research completes.