24 vertical TikTok videos — generated 100% locally

Everything runs offline: Manim for 2D animation, Blender (EEVEE, headless) for the 3D scenes, Godot (Jolt physics engine, headless) for rigid-body physics, qwentts.cpp (a cloned Rioplatense voice) for the narration and the keysound piano samples for the music.

The videos are in Spanish: narration, subtitles and on-screen text are Spanish on purpose. The code is in English.

The videos (`out/`)

File	Topic	Length	Hook
`montyhall.mp4`	The Monty Hall problem	39.5 s	sparks debate in the comments
`nueves.mp4`	0,999… = 1	38.8 s	sparks debate in the comments
`cumple.mp4`	The birthday paradox	38.6 s	surprising fact
`luna.mp4`	Folding paper to the Moon	38.4 s	surprising fact
`japones.mp4`	Multiplying with lines	38.8 s	useful, gets saved and shared
`benford.mp4`	Benford's law (spotting fraud)	42.8 s	surprising fact + useful
`hilbert.mp4`	The infinite hotel	37.1 s	sparks debate in the comments
`reuleaux.mp4`	Why manhole covers are round	36.0 s	satisfying to watch
`regla72.mp4`	The rule of 72	35.5 s	useful, gets saved
`buffon.mp4`	Getting π by dropping needles	36.7 s	mathematical beauty
`collatz.mp4`	The Collatz conjecture (3n+1)	46.2 s	unsolved mystery
`simpson.mp4`	Simpson's paradox	53.0 s	debate + useful
`cuerda.mp4`	A rope around the Earth	45.4 s	surprising fact
`gabriel.mp4`	Gabriel's horn	40.5 s	mathematical beauty
`mobius.mp4`	Möbius strip	46.3 s	try it at home
`bayes.mp4`	False positive (medical test)	46.9 s	debate + useful
`cicloide.mp4`	Brachistochrone: the fastest curve	47.6 s	satisfying to watch
`dados.mp4`	Non-transitive dice	50.6 s	sparks debate in the comments
`bolapeluda.mp4`	Hairy ball theorem (3D)	41.9 s	satisfying to watch
`caos.mp4`	Three double pendulums and 1 mm (3D)	40.6 s	hypnotic, watched to the end
`diferencial.mp4`	How a differential works (3D)	44.1 s	useful, gets saved and shared
`galton.mp4`	Galton board: 500 balls draw the bell curve (Blender + Godot)	45.7 s	satisfying + question in the comments
`bloques.mp4`	Harmonic stack: blocks overhanging the table (Blender + Godot)	48.0 s	try it at home + surprising fact
`domino.mp4`	Dominoes ×1.5: from 5 mm to the Aconcagua (Blender + Godot)	66.7 s	exponential + question in the comments (over 1 min)

Format: 1080×1920, 30 fps, H.264 + AAC 48 kHz stereo, −13.8 to −15.3 LUFS measured (TikTok's target is −14, so it neither lowers nor raises them), true peak below −1 dBTP. Burned-in subtitles: 80 % watch without sound.

Regenerating

conda run -n manim python scripts/gen_audio.py montyhall     # voice + check
conda run -n manim manim render -v ERROR --disable_caching \
    --format=mp4 -o montyhall scenes/v1_montyhall.py MontyHall
conda run -n manim python scripts/build_video.py montyhall   # music + mix + mux
conda run -n manim python scripts/verify_final.py montyhall  # QA on the final MP4

DRAFT=1 before rendering drops to 540×960 to iterate fast.

Layout

scripts/narration.py   scripts: (text that is spoken, on-screen subtitle)
scripts/gen_audio.py   synthesizes each sentence and checks it by transcribing with Whisper
scripts/numspell.py    turns "99,9%" into "noventa y nueve coma nueve por ciento"
                       so the transcription can be compared with the script
scripts/regen.py       regenerates one sentence trying 5 seeds
scripts/repick.py      picks the best take already synthesized again
scripts/verify.py      QA: transcribes every WAV and flags doubtful ones
scripts/verify_final.py  final QA: extracts the audio of the mixed MP4 and transcribes it,
                       which proves the voice is intelligible over the piano
scripts/music.py       piano bed (original progression) + ducking under the voice
scripts/build_video.py builds the track following the timeline and muxes it onto the video
scenes/tiktok.py       base: 9:16, subtitles, progress bar, "beats"
scenes/v*.py           one scene per video (Manim)
scenes/overlay.py      UI layer for the 3D videos: the same bar, chip and subtitles,
                       rendered with alpha
blender/base.py        base of the 3D scenes: palette, camera, primitives, timing
blender/*.py           one scene per 3D video
godot/sim.gd           generic rigid-body simulator fed by Blender
scripts/build_blender.py  composites background + 3D + overlay + audio in one pass
brand/                 channel logo and avatar (make_logo.py)

The 3D videos

Rendered headless with Blender 5.2 and EEVEE:

conda run -n manim manim -s --disable_caching -o fondo scenes/overlay.py Backdrop  # dot pattern
NAME=caos CHIP=Caos CHIPCOL=rosa conda run -n manim manim --transparent \
    --disable_caching -o caos scenes/overlay.py Overlay   # UI layer + timeline
/path/to/blender -b --factory-startup --python blender/caos.py  # PNG sequence with alpha
conda run -n manim python scripts/build_blender.py caos    # compositing + audio

Decisions that paid for themselves:

No keyframes. The render walks the frames one by one and calls refresh(f) before each one, which places everything. That way physics is computed exactly on every frame, skipped frames can be rendered to check the framing, and there is no fighting with Blender 5.x's layered actions API.
The background is not in the 3D scene. Blender renders with film_transparent and the dotted background is composited underneath with ffmpeg, so it is pixel for pixel the same as in the 18 Manim videos.
Metal needs something to reflect. studio_world() puts a horizon gradient in the world: it is not visible (the film is transparent) but it is the only thing that separates steel from grey plastic. Without it, a metallic material looks flat.

blender/base.py generates the mechanical parts in code:

Function	What it does
`bevel_gear`	Bevel gear with a cone half-angle and spiral teeth. Two mesh at 90° when their γ add up to 90: 45+45 for the differential ones, 72+18 for ring gear and pinion (and there the tooth ratio is fixed by sin72/sin18 = 3.07, which is why they are 43 and 14)
`full_wheel`	Tire with lugs and grooves + alloy rim with spokes + brake disc
`ring_sector`	Curved wall: two of these and two flanges make a differential housing with windows, which is what lets you see the gears inside
`flat_ring`, `screws`	Flanges and bolt circles

The intermediate PNGs stay in render/<name>/ (about 4 GB for three videos). They allow recompositing with other audio or subtitles without rendering again, which takes 30 to 55 minutes per video; if they are not needed, delete them.

SAMPLES=8 lowers the EEVEE samples (16 by default, 64 stock: that is 10 s/frame on a GTX 1050 and 1.6 s/frame with 8). EVERY=55 renders one frame out of every 55 to build a contact sheet without waiting for the whole render.

Sync is not estimated: the scene writes out/<name>_timeline.json with the real time of each block and the audio is placed at those exact marks.

Environment variables

Variable	What for
`PROJECT`	project root (defaults to the repository directory)
`MODE=sim`	in the Godot scenes: recompute the physics instead of rendering
`SAMPLES`, `PCT`	EEVEE samples and resolution percentage
`FROM_FRAME`, `TO_FRAME`, `EVERY`, `RESUME`	which frames to render
`QWENTTS_DIR`, `VOICE_REF`	qwentts.cpp and the reference voice for the narration
`PIANO_DIR`	keysound piano samples

The older Spanish names (PROY, MODO, MUESTRAS, DESDE, HASTA, SALTO, SEGUIR, PRUEBA) are still accepted.

Voice

The repository does not include the cloned voice. VOICE_REF is the path prefix of its three files (.spk, .rvq and .txt), extracted from a recording with qwen-codec --talker.

What is not in git

Everything generated: render/ (intermediate PNGs), media/ (Manim output), audio/ (narration), out/ (final videos and timelines), tmp/ and godot/work/ (simulations). They are regenerated with the steps above.

About the music

The progressions (Am–F–C–G, Em–C–G–D, Am–Em–Dm–E, Cm–G#–A#–Gm, A–E–F#m–D, etc.) are original and played with loose piano samples, so the audio is yours and Content ID will not flag it. If you prefer a trending TikTok sound, upload anyway and add it from the app: the voice is already mixed at −14 LUFS and stays intelligible over it.

License

MIT