Animation2Code (Image Frames) - Temporal Similarity: leaderboard
Video-to-code benchmark (Animation2Code, Ji, Mudunuri, Chan and Suhr, UC Berkeley, 2026): the model sees a CodePen web animation as image frames sampled at 2 FPS and must write one self-contained HTML/CSS/JavaScript file that reproduces it. The rendered result is compared with the reference inside the animated region by tracking points with CoTracker3 and scoring direction and speed agreement of the motion trajectories (Chamfer-averaged, at most 1). The score is the mean over the 214 test animations, zero-shot; higher is better.
Metric: Temporal similarity (-0.5 to 1; CoTracker3 tracklet direction and speed agreement, Chamfer-averaged; 214 test animations, zero-shot, video given as image frames sampled at 2 FPS). Source: arxiv.org. Saturation forecast: Around 2030. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 0.3 | #141 |
| 2 | Claude Sonnet 4.6 | 0.29 | #158 |
| 3 | GPT-5.4 | 0.29 | #121 |
| 4 | Qwen 3 VL 8B Instruct | 0.24 | #982 |
| 5 | Llama 4 Scout | 0.21 | #1787 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=animation2code-image-frames-temporal-similarity · How It Works · Data refreshed daily, snapshot 2026-11-12.