# benchmark.ai

> Leaderboards for AI image, video, world model and SVG generation, ranked by millions of pairwise human preference votes. Open data, versioned and citable.

Each benchmark ranks generative models on a fixed prompt set from pairwise human preference votes, fitted with a Bradley–Terry model and reported on the Elo scale with 95% confidence intervals. Every release is versioned and its votes are published on Hugging Face.

## Leaderboards

- [General image benchmark v1](https://www.benchmark.ai/image): Text-to-Image, judged on overall preference, coherence, prompt alignment. Leading now: GPT Image 2.5 (Sunburst), GPT Image 2.5 (Flare), GPT Image 2.0 (45 models ranked).
- [Camera movement benchmark v1](https://www.benchmark.ai/camera-movement): Camera Movement, judged on movement match. Leading now: Gemini Omni Flash 1.1, Wan 3.0, Happy Horse 1.1 (16 models ranked).
- [Relative camera movement benchmark v1](https://www.benchmark.ai/relative-camera-movement): Relative Camera Movement, judged on prompt alignment. Leading now: Gemini Omni Flash 1.1, Wan 3.0, Happy Horse 1.1 (16 models ranked).
- [World Model Physics benchmark v1](https://www.benchmark.ai/world-model-physics): Physics, judged on blind, with-reference. Leading now: FLUX 3 Max, Gemini Omni 1.1 Flash, MiniMax H3 (28 models ranked).
- [World Models Realism benchmark v1](https://www.benchmark.ai/world-models-realism): Realism, judged on photorealism, visual integrity. Leading now: MiniMax H3, Grok Imagine Video 1.5, Gemini Omni Flash 1.1 (18 models ranked).
- [Image-to-3D benchmark ](https://www.benchmark.ai/image-to-3d): Image-to-3D, judged on geometry, faithfulness. Leading now: Hi3D v3.0, Meshy v7.1, Meshy v7 (13 models ranked).
- [SVG generation benchmark v1](https://www.benchmark.ai/svg): SVG, judged on overall preference, coherence, prompt alignment. Leading now: quiver-arrow-2-telos, claude-opus-5.5-thinking, quiver-arrow-2 (69 models ranked).

## Methodology

- [Methodology](https://www.benchmark.ai/methodology): scoring, matchmaking, prompt curation and capability dimensions.
- [Annotator quality](https://www.benchmark.ai/annotator-quality): how human judgements are collected and filtered.
- [Updates](https://www.benchmark.ai/updates): releases, model additions and methodology changes.

## Optional

- [All benchmarks](https://www.benchmark.ai/benchmarks): live leaderboards and the rounds in preparation.
- [Datasets on Hugging Face](https://huggingface.co/rapidata): the published votes behind each release.