Code2Video—the specific function is exactly as the name implies: rendering animated videos using code.
Feed it some knowledge points or data, and you'll get "code-level" videos.
Current text-to-video creations are stunning, with gorgeous colors and smooth character movements...
However, text-to-video still struggles with certain scenarios that can't be solved yet, like precisely rendering data, which is hard to achieve. Once generated, it can't be edited; any adjustments require starting over. Also, if there's too much text or it's too complex, the effects aren't ideal.
The one below was generated using Veo, and it gets weird when touching on data.
Code2Video is tailored for scenarios demanding precise data; it doesn't replace everyday text-to-video tools.
It's not limited to tutorial videos—many science communicators (like Xiao Lin Shuo, have you watched it?) produce flashy animations to demonstrate data changes. Before, this meant painstaking frame-by-frame rendering with extremely high barriers and requiring tons of skills. Now, this tool makes it direct and simple.
She mentioned that making one video takes about half a month; AI will undoubtedly boost efficiency exponentially.
If these videos are produced with higher quality, they're incredibly powerful. I'm tempted to start such an account myself.
Science popularization and educational videos all greatly benefit from Code2Video.
Project Introduction
Code2Video is an open-source project from the National University of Singapore that automatically generates high-quality educational videos via code-driven animations. The output videos are highly controllable, reproducible, and editable.
DEMO
After watching the DEMO, take a look at those science short videos—you'll understand just how useful Code2Video is.
Key Features
1. Code-driven, offering greater control.
It automatically writes Manim Python code to manage graphics, layouts, and animation rhythms, resulting in clearer video structures and more precise visuals.
2. Three-agent collaboration mechanism for a fully automated closed loop.
• Planner: Understands knowledge points and plans video content. Given a knowledge point, it automatically designs the teaching sequence—what comes first, what follows, and which graphics or animations to display at each step.
• Coder: Based on the Planner's guidance, it uses AI to generate runnable Manim animation code—the true script of the video.
• Critic: Once the animation is rendered, it assesses clarity of visuals, layout reasonableness, and smoothness of animations, then provides feedback to the Coder.
3. Error correction and code self-tuning mechanism
The Coder module features comprehensive auto-debugging. When Manim code errors occur, it proactively identifies and attempts fixes. Post-rendering, it adjusts layouts, animation pacing, and element positions based on Critic's visual feedback via multiple iterations for optimal results.
In essence, the system auto-fixes code → re-renders → re-checks,
looping until the video effect is satisfactory.
Project Link