Launching today
explainer-video
Turn any paper into a 3Blue1Brown-style explainer video
2 followers
Turn any paper into a 3Blue1Brown-style explainer video
2 followers
Sometimes it is super helpful to watch a video or animation to understand something technical before going into the formulas / algorithms, etc. So I wrote a skill that (completely locally) generates a video for you. You ask your agent for a video and it writes the animation code, prompts a small local model (LFM3-audio) to do the voice, and you get an MP4 explaining what you need to understand. Completely open source! Just check the github (linked) to install it :)
Free
Launch Team / Built With
Dropbox Sign APIBuild eSignatures into your app with less work.
Promoted
Hey PH! I'm Manu, an MS student at Stanford. I fine-tune small models and build AI for learning :)
tl;dr: Explainer Video is an agent skill that turns a paper, a repo or any topic into a narrated, 3Blue1Brown-style video. One command to install: npx skills add manoletre/explainer-video or even better: ask your agent to install the skill.
The skill is completely open source and free so feel free to edit it to match your needs!
why I built this: Sometimes an image or a clear animation is better than a thousand words. When reading a paper, going over the math is important, but a short video beforehand can really help. But good explainer videos only exist for the famous stuff.
what it does: You ask your agent something like "make an explainer video of this paper and how it applies to this project" and it:
reads the source and writes a cheatsheet
outlines 8-13 sections (~6-10 min)
writes the narration + a ManimGX scene (thanks academa.ai), with animations timed to specific words
speaks every line with LFM3-Audio from Liquid AI, locally through llama.cpp (no TTS API, no keys)
transcribes each take back with the same model and keeps the one with the lowest word error rate. Anything above 15% gets flagged so the agent fixes it
checks storyboard frames + a layout report, then renders the MP4 + .srt subtitles
On my M3-series Mac an 8-minute video takes about 1-2 mins to generate end to end. There are two example videos in the repo (data structures + search algorithms for interview prep).
honest limits:
setup is ~3 GB and ~10 min (ffmpeg, uv, ManimGX, the voice model)
macOS and Linux only for now
the voice is a 2B model. It's really good for its size, but it's not ElevenLabs. Given that its open source you could change the voice to use something like ElevenLabs, but I did not want to pay for every video, just use my current claude subscription and a local model
what's next: a video where the skill explains itself (meta, I know), more styles than dark-3b1b, picking your own voice, etc. But honestly this is opensource and free so feel free to improve it, fork it, suggest PRs, etc :)
If you want to see what comes next (and the videos people request), follow me on X: @manoletre. And if you try it, a star on GitHub helps a lot!
Big thanks to Academa for ManimGX and to the Liquid AI team for LFM3-Audio <3