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

A new benchmark called Animation2Code aims to test whether vision-language models can reconstruct executable code from videos, moving beyond static image-to-code tasks to evaluate temporal visual reasoning. The benchmark, introduced in a paper submitted to arXiv on June 26, 2026, contains 1,069 web animation videos paired with their underlying HTML, CSS, and JavaScript implementations [1][2]. The dataset spans diverse visual appearances and motion patterns, providing a controlled testbed for assessing how well models capture dynamic content [2]. Recent vision-language models have posted strong results on static visual-to-code problems, such as generating code for webpages, charts, or SVGs [1][2]. However, the authors note that it remains unclear whether these systems can recover temporal dynamics when motion is involved [2]. Animation2Code was designed to fill that gap by requiring models to produce executable animation code from video inputs [1]. Two human-aligned evaluation metrics accompany the dataset: appearance similarity and temporal similarity [1][2]. These metrics are intended to separate visual fidelity from temporal alignment when comparing rendered animations against ground-truth samples [2]. In benchmarking experiments, state-of-the-art VLMs struggled to maintain temporal consistency in their reconstructions, even when they achieved high scores on appearance similarity [1][2]. The performance gap persisted under finetuning and iterative refinement settings [2]. The paper appears on arXiv, an open-access repository that hosts electronic preprints across fields including computer science, physics, and mathematics [6]. As of November 2024, the repository receives roughly 24,000 submissions per month and has surpassed two million total articles [6]. The Animation2Code submission falls under the Computer Vision and Pattern Recognition category [1]. Code and data for the benchmark are publicly available through a project website linked in the paper [1][2]. The authors frame the work as a step toward understanding whether current multimodal models can reason about time-dependent visual information, a capability that static benchmarks do not measure [2].