Skip to content

Roadmap

manimvox lets an AI agent make high-quality educational videos at low cost, with a person directing. Each release below moves one step closer to that goal. Items are planned, not promised; order and scope may change. Discuss any of them in GitHub issues.

Released: 0.2.0 (alpha), on PyPI, tested on Linux, macOS and Windows.

People should see what manimvox makes before they install it, and be able to compare the choices it offers.

  • [x] Documentation site, published from this repository with GitHub Pages: install, quickstart, authoring, voices, blocks, 3D, agent skills, and the Python API from docstrings.
  • [x] Gallery of finished videos, each with its source project and the command that made it, opening with the heart explainer (a CT heart cut open, blood flow).
  • [x] Block gallery, generated from each block's preview and demo.
  • [x] A short demo (GIF or video) at the top of the README.
  • [x] Benchmarks of harnesses, models and voices, with the method and every record published: first, Sarvam AI, Gnani.ai and ElevenLabs on an episode made by Claude Code with Sonnet 5.5.
  • [x] A leaderboard on each gallery example with benchmark runs: Voice, Harness and Model tables generated from the runs' records, so a new run updates every page it belongs to.
  • [x] Voices for Indian languages: Sarvam AI and Gnani.ai.
  • [x] Fixes from the 0.1.0 showcase: the Kokoro status line ends its sentence; skills install falls back to the built-in installer on Node older than 22.20; scripts/release.py sets the version everywhere at once.

Next

  • [x] Even loudness across voices. Every narration line is levelled to one loudness (−16 LUFS by default), so changing the voice doesn't change the volume. The voice benchmark's 7.9 LU spread is now 0.1.
  • [ ] More gallery videos: a paper video next, then a concept video and a PR walkthrough.
  • [ ] More benchmark runs: other harnesses (Codex, Cursor, Gemini CLI) and models on the same prompts, and more voices, including Hindi narration. Runs are made by hand for now: each costs model and voice API fees.

0.3: professor-quality visuals

Make the default output look like a good lecture, not slides. The 0.1.0 heart explainer only got there with pieces written inside the episode; they become part of the kit.

  • [ ] Leader-line labels. A dot on the part, a thin leader line, and the name in a column beside it, laid out so labels don't collide.
  • [ ] Flow streams. Glowing particles along a route that fade in where they enter and out where they leave, and can be dimmed: blood, data, current, traffic.
  • [ ] Entry and exit markers. Dashed arrows for "comes in here" or "leaves the picture here".
  • [ ] Calmer kit defaults. Fewer boxes, thinner lines, more space.
  • [ ] A visual-quality checklist in the skills, which agents run before a final render.

0.4: 3D and real data

Real objects and real data, used with their licences respected.

  • [ ] manimvox assets. Fetch a model or scan into research/, after asking, and record its source, licence and credit line in FACTS.md. Refuse a licence that doesn't allow the use.
  • [ ] Cut-open view block. Cut a volume mesh (a scan, a simulation) along a plane, colour its regions, and give the points to label. This generalises the heart explainer's CT cut.
  • [ ] Better mesh-model shading. manimvox's own lighting instead of Manim's wash towards white, and faster drawing of large models.

Later

  • Faster renders: scenes in parallel, smarter caching.
  • Narration and subtitles in more languages.
  • Automated evals: every gallery submission rendered and scored by the benchmark, with its leaderboard published alongside. Until then, runs are made by hand, because each run costs model and voice API fees.