manim-studio

Prompt → OpenAI writes a Manim scene → renders on InstaVM → MP4.

Live: https://singing-debonair-goose.instavm.site/

Architecture

• FastAPI app on a long-lived InstaVM VM (manimstudio/app.py).
• Each render spawns a short-lived VM from a pre-built base image (manim==0.20.1 + cairo/pango/ffmpeg). Cold start ~4s.
• Long-poll job API (/api/jobs, /api/jobs/{id}/poll).

Files

manimstudio/
  app.py            FastAPI app + long-poll endpoints
  generator.py      OpenAI prompt → manim scene → render on InstaVM
  db.py             SQLite job store
  static/           UI
scripts/
  deploy.py         Provision app VM, push code, start systemd unit
  setup_vault.py    Store OpenAI key in InstaVM vault
docker/
  Dockerfile        Public manim base image (instavm/manim-base:0.20.1)
.snapshot_id        Optional: private snapshot UUID (gitignored)

Setup

Get InstaVM key: https://dashboard.instavm.io and OpenAI key: https://platform.openai.com

bash
pip install -U instavm openai fastapi uvicorn
export INSTAVM_API_KEY=...
export OPENAI_API_KEY=...

# one-time: store openai key in instavm vault
python scripts/setup_vault.py

# deploy (uses the public manim base image by default)
python scripts/deploy.py

Tuning

• Render VM: memory_mb=2048, vcpu_count=2 in generator.py.
• Base image: MANIM_BASE_OCI_IMAGE (default instavm/manim-base:0.20.1).

Known gotchas

• deploy.py::find_or_create_vm checks status in ("running","starting","ready") but the API returns "active" — so each deploy spawns a fresh VM. Cosmetic.
• Share subdomain is randomly assigned by InstaVM.