Animation2Code: Evaluating Temporal Visual Reasoning in Video-to-Code Generation

Animation2Code is a benchmark of 1,069 short web animations paired with the self-contained HTML documents that produce them. Each example is sourced from a publicly published CodePen pen, normalized to plain HTML/CSS/JavaScript, and rendered to MP4 with a headless Chromium recorder. The benchmark is split into 769 training, 86 validation, and 214 test examples. The intended use is evaluating whether multimodal language models can recover the symbolic program (HTML/CSS/JavaScript) that produces a given short web animation video, and training models on the same task. Each example directory contains animation.mp4 (the rendered video, 1024x768 at 30 FPS, 2-8 seconds), full.html (the self-contained HTML document), and metadata.json (title, original CodePen URL for attribution, and external resource dependencies). Released under CC BY 4.0; see README.md and croissant.json for full documentation, license terms, and Responsible AI metadata.