Head-to-head human preference ranking for text-to-video, image-to-video, and video-edit models.
Video Arena is the video-generation companion to the Arena.ai chat leaderboard. Users see two anonymous generated clips for the same prompt and vote on which they prefer. The pairwise wins drive a Bradley-Terry rating that reflects general taste rather than any single technical dimension. Arena.ai now runs three separate video boards: text-to-video at arena.ai/leaderboard/text-to-video, image-to-video at arena.ai/leaderboard/image-to-video, and video-edit at arena.ai/leaderboard/video-edit. We report the text-to-video rating on this page.
Voters do not see which model produced which clip. Aggregate wins feed a Bradley-Terry rating per model. We normalize the published rating to a 0–100 scale here so it can be read alongside VBench. The text-to-video board scores prompt-to-clip generation; image-to-video scores motion conditioned on an input image; video-edit scores edits to an input clip with a text instruction.
| # | Model | Lab | Source | Score |
|---|---|---|---|---|
| 01 | Seedance 2.0 | ByteDance | Closed | 100.0 |
| 02 | Runway Gen-4.5 | Runway | Closed | 82.1 |
| 03 | Veo 3 | Closed | 81.4 | |
| 04 | Kling 2.5 Turbo | Kuaishou | Closed | 78.1 |
| 05 | Veo 3.1 | Closed | 76.6 | |
| 06 | Kandinsky 5.0 Video Pro | Kandinsky | Open | 64.2 |
| 07 | LTX-2 19B | Lightricks | Open | 49.3 |
| 08 | Kandinsky 5.0 Video Lite | Kandinsky | Open | 42.0 |
| 09 | Wan2.2-T2V-A14B | Alibaba | Open | 42.0 |
| 10 | HunyuanVideo-1.5 | Tencent | Open | 10.6 |
| 11 | Mochi 1 Preview | Genmo AI | Open | 0.0 |
6 model(s) with undisclosed parameter counts not shown. Most closed-source labs do not publish model size.
Video Arena measures general human preference. VBench breaks the task into 16 specific quality dimensions and reports each one. Use both together: VBench tells you where the weaknesses are; Arena tells you whether users care.
Based on score correlations across our database.
