anim2code/dataset

Animation2Code

Animation2Code is a benchmark for video-to-code generation of web animations. It contains 1,069 examples, each with a rendered web animation video and self-contained HTML/CSS/JavaScript that produced it.

• Paper: Animation2Code: Evaluating Temporal Visual Reasoning in Video-to-Code Generation (NeurIPS 2026)
• Project page: https://anya-ji.github.io/animation2code-website/

Dataset Structure

data/
├── train/   769 examples
├── val/      86 examples
└── test/    214 examples
    └── codepen-<collection_id>-<pen_id>[-<k>]/
        ├── animation.mp4
        ├── full.html
        └── metadata.json

Each example directory contains exactly three files:

File	Description
animation.mp4	The rendered reference video (the task input). H.264 (yuv444p), 1024×768, 30 fps, about 2–8 s.
full.html	A self-contained HTML document with inline <style> and <script> tags (the ground-truth target).
metadata.json	Provenance and dependency information (see below).

Example IDs

codepen-<collection_id>-<pen_id> names the CodePen collection and pen the example came from. When a single pen contained several independent animations, it was split into sub-examples with a -<k> suffix (e.g. codepen-DrPkOq-poyOMgr-1 … -4). All sub-examples of a pen are in the same split.

Metadata fields

In metadata.json:

Field	Presence	Description
title	always	Pen title given by the original author.
url	always	Canonical CodePen URL, kept for attribution.
external_resources	486 / 1,069	{"scripts": [...], "stylesheets": [...]}: URLs of external libraries and stylesheets the pen loads. These are not inlined into full.html and must be loaded from the network when rendering.
jquery_translated	41 / 1,069	The jQuery URL that was removed when the pen's jQuery code was translated to vanilla JavaScript.

Example:

{
  "title": "Merry Christmas Tree!",
  "url": "https://codepen.io/chrisgannon/pen/dypvKvR",
  "external_resources": {
    "scripts": [
      "https://unpkg.co/gsap@3/dist/gsap.min.js",
      "https://s3-us-west-2.amazonaws.com/s.cdpn.io/16327/MorphSVGPlugin3.min.js"
    ]
  }
}

Usage

The data is stored as raw files (one directory per example), so download it with huggingface_hub:

import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download("anim2code/dataset", repo_type="dataset"))
# For a single split only: snapshot_download(..., allow_patterns="data/test/*")

for ex_dir in sorted((root / "data" / "test").iterdir()):
    video_path = ex_dir / "animation.mp4"
    html = (ex_dir / "full.html").read_text()
    meta = json.loads((ex_dir / "metadata.json").read_text())

Or with the CLI:

hf download anim2code/dataset --repo-type dataset --local-dir anim2code_data

Dataset Statistics

Split	Examples
train	769
val	86
test	214
total	1,069

Durations: 2–8 s (mean 5.1 ± 2.4 s). Looping animations are captured for at least one full cycle.

License

This dataset is released under CC BY 4.0. The original code comes from public CodePen pens, which are MIT-licensed (CodePen licensing). Each example's metadata.json keeps the original pen URL so the original author can be credited.

Citation

@article{ji2026animation2code,
  title         = {Animation2Code: Evaluating Temporal Visual Reasoning in Video-to-Code Generation},
  author        = {Ji, Anya and Mudunuri, Abhijith Varma and Chan, David M. and Suhr, Alane},
  year          = {2026},
  eprint        = {2606.28593},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2606.28593},
}