RayMorTwinkle/anytomanim

AnyToManim — CLI toolchain for narrated 3b1b-style videos

The product is a CLI, not an API harness. You (the host agent) do the creative stages — write the brief, the beat sheet, the storyboard, the scene code — guided by the spec files shipped in the package. The CLI does everything deterministic: TTS, duration measurement, rendering, concatenation.

Install & check

pipx install anytomanim          # or: pip install -e . in a venv
anytomanim doctor                # reports missing deps + fix hints

Workflow — export the spec, follow it

anytomanim export-guides ./guides   # WORKFLOW/DIGEST/SCRIPT/SCENE_GUIDE + schema

Then follow guides/WORKFLOW.md:

1. Write brief.md per DIGEST.md
2. Write beats.json per SCRIPT.md (+ beats-schema.json). Read SCRIPT.md first: narration has to chain, not list.
3. Read ORCHESTRATION.md, then write storyboard.json per STORYBOARD.md. Open the runnable recipes for the cards you pick. Run anytomanim storyboard storyboard.json --beats beats.json and fix it.
4. anytomanim tts --beats beats.json — synthesizes every beat, writes audio_file + real audio_seconds back (re-run after edits: only changed beats)
5. Write scenes/*.py per SCENE_GUIDE.md (NarratedScene + MiMoService(beats_file=...), with self.voiceover(beat="b01") as tr). Follow the storyboard; start from the matching recipes/.
6. anytomanim render scenes/X.py --quality ql — fix & retry on error
7. anytomanim preview <mp4> — READ sheet.png and the per-beat report (overlaps, overruns, frozen beats, small content, empty bands, a fixed header, stage changes the storyboard does not mention); fix and re-render until clean
8. Render --quality qh, anytomanim assemble a.mp4 b.mp4 -o final.mp4, anytomanim subtitles beats.json --mux final.mp4

Iterate without a key / without cost: ANYTOMANIM_TTS=draft (local voice).

anytomanim run "<topic>" also exists — an embedded-LLM reference driver for automated batch eval (requires an LLM API key; interactive agents should use the guide-driven flow instead).

Contracts

• beats.json is THE interface — public IR (narration + visual + audio_seconds).
• Audio-first sync: synthesize → measure → every run_time derives from tr.duration, and the render replays exactly those wavs. Never invent durations.
• A clean render is not done — look at the preview sheet every time.
• Continuity comes first. Before an object leaves, ask whether it can become the next thing. Fade out and redraw is the last resort.
• Draw the relation when you can, instead of writing it as a label.
• Voices: --voice baihua|moli|milo; clone your own voice via assets/voice/<name>.mp3 + voices.json entry. zh + en supported.
• See examples/ in the repo for golden end-to-end runs.