Code2Video: A Code-centric Paradigm for Educational Video Generation

Yanzhe Chen, Kevin Qinghong Lin, Mike Zheng Shou
Show Lab @ National University of Singapore

  📄 Paper   |     🤗 Daily Paper   |     🤗 Dataset   |     🌐 Project Website   |     💬 X (Twitter)

Learning Topic	Veo3	Wan2.2	Code2Video (Ours)
Hanoi Problem
Large Language Model
Pure Fourier Series

• [2025.10.6] We have updated the ground truth human-made videos and metadata for the MMMC dataset.
• [2025.10.3] Thanks @_akhaliq for sharing our work on Twitter!
• [2025.10.2] We release the arXiv, code and dataset .
• [2025.9.22] Code2Video has been accepted to the Deep Learning for Code (DL4C) Workshop at NeurIPS 2025.

• 🌟 Overview
• 🚀 Quick Start: Code2Video
  • 1. Requirements
  • 2. Configure LLM API Keys
  • 3. Run Agents
  • 4. Project Organization
• 📊 Evaluation: MMMC
• 🙏 Acknowledgements
• 📌 Citation

Code2Video is an agentic, code-centric framework that generates high-quality educational videos from knowledge points.
Unlike pixel-based text-to-video models, our approach leverages executable Manim code to ensure clarity, coherence, and reproducibility.

Key Features:

• 🎬 Code-Centric Paradigm — executable code as the unified medium for both temporal sequencing and spatial organization of educational videos.
• 🤖 Modular Tri-Agent Design — Planner (storyboard expansion), Coder (debuggable code synthesis), and Critic (layout refinement with anchors) work together for structured generation.
• 📚 MMMC Benchmark — the first benchmark for code-driven video generation, covering 117 curated learning topics inspired by 3Blue1Brown, spanning diverse areas.
• 🧪 Multi-Dimensional Evaluation — systematic assessment on efficiency, aesthetics, and end-to-end knowledge transfer.

cd src/
pip install -r requirements.txt

Here is the official installation guide for Manim Community v0.19.0, to help everyone correctly set up the environment.

Fill in your API credentials in api_config.json.

• LLM API:

  • Required for Planner & Coder.
  • Best Manim code quality achieved with Claude-4-Opus.

• VLM API:

  • Required for Planner Critic.
  • For layout and aesthetics optimization, provide Gemini API key.
  • Best quality achieved with gemini-2.5-pro-preview-05-06.

• Visual Assets API:

  • To enrich videos with icons, set ICONFINDER_API_KEY from IconFinder.

We provide two shell scripts for different generation modes:

Script: run_agent_single.sh

Generates a video from a single knowledge point specified in the script.

sh run_agent_single.sh --knowledge_point "Linear transformations and matrices"

Important parameters inside run_agent_single.sh:

• API: specify which LLM to use.
• FOLDER_PREFIX: output folder prefix (e.g., TEST-single).
• KNOWLEDGE_POINT: target concept, e.g. "Linear transformations and matrices".

Script: run_agent.sh

Runs all (or a subset of) learning topics defined in long_video_topics_list.json.

sh run_agent.sh

Important parameters inside run_agent.sh:

• API: specify which LLM to use.
• FOLDER_PREFIX: name prefix for saving output folders (e.g., TEST-LIST).
• MAX_CONCEPTS: number of concepts to include (-1 means all).
• PARALLEL_GROUP_NUM: number of groups to run in parallel.

A suggested directory structure:

src/
│── agent.py
│── run_agent.sh
│── run_agent_single.sh
│── api_config.json
│── ...
│
├── assets/
│   ├── icons/          #  downloaded visual assets cache via IconFinder API
│   └── reference/      # reference images
│
├── json_files/         # JSON-based topic lists & metadata
├── prompts/            # prompt templates for LLM calls
├── CASES/              # generated cases, organized by FOLDER_PREFIX
│   └── TEST-LIST/      # example multi-topic generation results
│   └── TEST-single/    # example single-topic generation results

We evaluate along three complementary dimensions:

1. Knowledge Transfer (TeachQuiz)

   python3 eval_TQ.py

2. Aesthetic & Structural Quality (AES)

   python3 eval_AES.py

3. Efficiency Metrics (During Creating)

   • Token usage
   • Execution time

👉 More data and evaluation scripts are available at: HuggingFace: MMMC Benchmark

• Video data is sourced from the 3Blue1Brown official lessons. These videos represent the upper bound of clarity and aesthetics in educational video design and inform our evaluation metrics.
• We thank all the Show Lab @ NUS members for support!
• This project builds upon open-source contributions from Manim Community and the broader AI research ecosystem.
• High-quality visual assets (icons) are provided by IconFinder and Icons8, which were used to enrich the educational videos.

If you find our work useful, please cite:

@misc{code2video,
      title={Code2Video: A Code-centric Paradigm for Educational Video Generation}, 
      author={Yanzhe Chen and Kevin Qinghong Lin and Mike Zheng Shou},
      year={2025},
      eprint={2510.01174},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.01174}, 
}

If you like our project, please give us a star ⭐ on GitHub for the latest update.