Video Generation Benchmarks¶
Artificial intelligence (AI) video generation is ranked the same way as image generation: blind human-preference arenas, where voters compare two clips generated from the same prompt and an Elo-style rating accumulates (the image page has a diagram of the mechanism). Video leaderboards also split into separate tasks: text-to-video (a clip from a written prompt), image-to-video (animating a supplied still while preserving its composition), and increasingly video editing (instruction-based changes to an existing clip). A model can lead one task and trail badly on another, so check the task that matches your use before comparing ranks.
What separates video models in practice, and what to look past the overall rank for:
- Temporal consistency: objects keep their identity across frames.
- Motion plausibility: physics that does not distract.
- Prompt adherence over time: events happen in the requested order.
- Clip length: how long a clip runs before quality decays.
Many video models now generate sound as well. Some leaderboards rank clips with and without audio in separate pools, so check which one you are reading.
The short narrated film How to Read a Leaderboard, on the Language Model Benchmarks page, covers image and video arenas as well.
The Leaderboards Worth Knowing¶
-
Artificial Analysis: Text-to-Video
Blind-vote arena rankings for generating clips from written prompts, proprietary and open-weights models together.
-
Artificial Analysis: Image-to-Video
Rankings for animating a supplied still image, scored with and without audio.
-
Arena: Text-to-Video
The Arena team's (formerly LMArena) blind-vote video arena, a useful second opinion on the same question.
-
Arena: Image-to-Video
Arena's image-animation rankings, from blind votes on two animations of the same still.
-
Artificial Analysis: Video Editing
Rankings for changing an existing clip by instruction, rated separately from generating one.
-
Arena: Video Edit
Arena's rankings for instruction-based edits to an existing clip, from blind head-to-head votes.
-
Artificial Analysis: Video Models
The companion view: generation time and price per minute of video beside the quality ratings, with charts of quality against each.
A Note on Medical Use¶
Everything on the image generation page about medical contexts applies with more force here: there is no public leaderboard for clinically accurate generated video, preference scores say nothing about anatomical or procedural correctness, and a fluent clip of a procedure can be confidently wrong in ways a non-expert will not catch. It is worth checking a generated clip for anatomical and procedural accuracy, yourself or with a colleague who knows the procedure, before using it in teaching.
A generated clip can show a real, identifiable person saying or doing things they never did. Presented as genuine, it misleads the audience and misrepresents the person. Asking them first and labeling the clip as generated lets viewers tell it from a real recording, and respects someone whose face and voice are being used.
Many of the tools being ranked are listed under Video Generation in the directory; for text models, see the Language Model Benchmarks.
Content last reviewed October 12, 2026. Review dates are maintained as described on the About page.