How Code2Video Turns Code into High‑Quality Teaching Videos in Minutes

Overview

Code2Video is an open‑source, code‑driven video generation framework that converts a textual knowledge point (e.g., “matrix and linear transformations”) into an executable Manim script, renders animated scenes, adds subtitles and illustrations, and outputs a polished teaching video.

Core Architecture

Three AI agents :

Planner – decomposes the knowledge point into a storyboard (introduction, principle, example, summary) to ensure logical flow.

Coder – generates Manim code that defines precise animation logic (e.g., vector rotation, tower of Hanoi disks) and inserts subtitles and graphics.

Critic – evaluates visual layout, color scheme, and overall aesthetic, then refines the script.

Dataset : A built‑in MMMC benchmark provides 117 high‑frequency topics ranging from linear algebra to algorithms, enabling out‑of‑the‑box video generation without manual design.

Reproducibility : The generated Manim code is fully open‑source and can be edited (e.g., change animation speed, number of disks) to produce variant videos while preserving style consistency.

Evaluation framework : Scores each video on knowledge transmission, visual appeal, and generation efficiency, and can automatically fix code bugs.

License : Distributed under the MIT license, allowing unrestricted local deployment.

Typical Use Cases

1. Teachers creating mathematics videos

Install Python 3.8+ and Manim Community v0.19.0.

Run the single‑concept script with the desired knowledge point:

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

The Planner creates a storyboard, the Coder produces Manim code for vector rotation and scaling, and the Critic optimizes layout.

Render the video (≈10 minutes) and deliver a clear animation that illustrates matrix effects.

Modify parameters in the generated script (e.g., number of disks in a Tower of Hanoi animation) to create new variants in minutes.

2. Technical bloggers producing algorithm series

Add topics such as "Quick Sort", "Binary Search", "Tower of Hanoi" to json_files/long_video_topics_list.json.

Run the batch script: sh run_agent.sh --MAX_CONCEPTS 5 The three agents automatically generate videos with consistent styling, subtitles, and algorithmic visualizations.

New topics can be added simply by editing the JSON file; no redesign is required.

3. Students creating personal revision videos

Specify a concept (e.g., "Pure Fourier Series") and select an academic visual style.

The system explains the definition, then animates the decomposition of a square wave into sine components.

Exported Manim code allows the student to adjust parameters such as the number of sine waves.

Getting Started

Step 1 – Prepare the environment

Ensure Python 3.8+ is installed and install Manim Community v0.19.0.

Clone the repository and install Python dependencies:

git clone https://github.com/showlab/Code2Video.git
cd Code2Video/src
pip install -r requirements.txt

Step 2 – Configure API keys

Edit api_config.json to provide required keys:

LLM API (recommended Claude‑4‑Opus) for high‑quality Manim code generation.

VLM API (recommended Gemini‑2.5‑Pro) for layout optimization.

Optional IconFinder API for additional icons.

Step 3 – Generate videos

Single‑concept mode (new users) :

sh run_agent_single.sh --knowledge_point "Tower of Hanoi"

Output is saved in CASES/TEST-single/.

Batch mode (series production) : edit json_files/long_video_topics_list.json with multiple concepts and run:

sh run_agent.sh --MAX_CONCEPTS 3  # generate the first three videos

Repository

Project source code:

https://github.com/showlab/Code2Video