animation2code benchmark
For best compatibility, please view this dashboard in a Chrome browser.

Zero-shot video (or image-frame) → code results on the test set, across commercial and open-source models.

Each output is tagged with A = appearance similarity and T = temporal similarity; higher is better for both. Click a video to inspect its code.

4956 of 214

ground truth

Only CSS: Milk

model outputs

GPT-5.4

A 0.87T 0.37

Claude Sonnet 4.6

A 0.84T 0.33

LLaMA 4 Scout

A 0.57T 0.04

ground truth

Only CSS: Kirby

model outputs

GPT-5.4

A 0.94T 0.27

Claude Sonnet 4.6

A 0.90T 0.31
no output

LLaMA 4 Scout

A T

ground truth

Only CSS: Bicycle

model outputs

GPT-5.4

A 0.85T 0.36

Claude Sonnet 4.6

A 0.83T 0.25

LLaMA 4 Scout

A 0.41T 0.29

ground truth

Only CSS: Screw 2

model outputs

GPT-5.4

A 0.70T 0.26

Claude Sonnet 4.6

A 0.57T 0.19

LLaMA 4 Scout

A 0.42T 0.00

model outputs

no output

Qwen3-VL-8B-Instruct

A T

GPT-5.4

A 0.77T 0.16

Claude Sonnet 4.6

A 0.80T 0.33

LLaMA 4 Scout

A 0.36T 0.25

model outputs

GPT-5.4

A 0.69T 0.39

Claude Sonnet 4.6

A 0.67T 0.43

LLaMA 4 Scout

A 0.44T 0.00

model outputs

GPT-5.4

A 0.72T 0.57

Claude Sonnet 4.6

A 0.75T 0.42

LLaMA 4 Scout

A 0.50T 0.49

model outputs

GPT-5.4

A 0.69T 0.24

Claude Sonnet 4.6

A 0.64T 0.25

LLaMA 4 Scout

A 0.50T 0.22