Booting · 00:00:00

Tom LEFRERE · Data Scientist

Raw data. A signal.

0%
EN FR

← Portfolio

· llm · python

Manim Animations

Pedagogical animations generated with Manim driven by an LLM, to visually explain technical concepts like RAG.

Manim Animations

The project

This is something I care about: I love explaining things visually, but making animations takes forever. Discovering Manim (the library behind 3Blue1Brown’s videos) opened an opportunity: use an LLM to quickly generate coherent visual representations of technical concepts. Here it was RAG, but obviously the approach works for any technical subject, especially the ones we explain poorly in person.

My contributions

LLM-driven Manim scripts to generate technical animations. Two variants of RAG animation (continuous and smooth) with different visual approaches. Optimized the LLM-to-Manim-to-GIF workflow to save a bit of iteration time.

Takeaways

Manim is an incredible tool for technical popularization. Pairing an LLM with an animation-generation tool opens huge possibilities to produce pedagogical content quickly. The main challenge is still getting clean visuals on the first try, which doesn’t always happen, and you often need to re-prompt a bit.

Context

Personal project, born of the urge to popularize complex AI concepts (RAG here) visually and elegantly, without spending days on it. The idea was really to have a fast production pipeline for animations.

A one-shot example

Cosine similarity, animation generated from a single prompt then rendered via Manim.

Tech stack

  • Python

  • Manim

  • LLM (assisted generation)

Cover image generated by Nano Banana (Google), since I no longer have the mockup template I used before. So it’s not a screenshot of the actual tool.