Vision-language models can turn a screenshot into working HTML, but hand them a video and ask them to reproduce the motion, and they fall apart.
Researchers introduced Animation2Code, a benchmark built from 1,069 web animation videos paired with their original HTML, CSS, and JavaScript source code. The dataset covers a range of visual styles and motion patterns. Two new metrics - appearance similarity and temporal similarity - let evaluators separate how well a model captures the look of an animation from how well it captures the movement. When the team ran state-of-the-art vision-language models against this benchmark, the results were consistent: models could often approximate the visual appearance but failed to maintain temporal consistency. Finetuning and iterative refinement did not fix the gap.
The finding matters because the field has used static tasks - webpage screenshots, charts, SVGs - as the primary measure of visual-to-code ability. Animation2Code is an argument that those benchmarks are too easy: they test pattern matching, not reasoning about time. A model that scores well on static tasks but cannot reconstruct motion has a fundamental gap, not a polish problem.
For anyone building AI coding tools that touch interactive or animated UI, this is a useful reality check - the benchmark is public, and the gap it documents is large enough that closing it will require more than a larger dataset or another fine-tuning run.