00:00:00:00 01 / 05 · HERO

01 — OPEN SOURCE · LOCAL FIRST · 9:16

AI video
pipelines
that don't
look like AI.

Turning "hidden cafes in Tokyo" into a reel.

A CLI that takes a text topic and returns a color-graded, captioned 9:16 reel. CLIP semantic footage matching, real color science, word-synced captions. Runs locally. Costs pennies.

View on GitHub Watch samples

reelgen — zsh

runs locally · no cloud · bring your own key

9:16 · 1080×1920

● REC

00:00:14:23

hidden cafes in tokyo

zero human edits

02 / 05 PIPELINE 00:01:28:00

TEXT TOPIC → VERTICAL REEL

topic → reel. seven stages.

scroll to trace the pipeline

INPUT

STORY

CLIP

GRADE

MOTION

CAPS

REEL

Text input

topic · mood · style — CLI flags

Two hooks generated per topic. Best semantic footage match wins.

TEXT TOPIC → VERTICAL REEL

topic → reel.
seven stages.

01

Text input

topic · mood · style flags — CLI args

02

LLM storyboard

two hooks generated — best match wins

03

Semantic clip match

CLIP scoring — 4-signal QA

04

Color grade

FFmpeg color science, not filters

05

Pacing + motion

cut timing · human-feeling intervals

06

Word-synced captions

Whisper · frame-accurate burn

07

Final reel

reel.mp4 ✓ $0.01

03 / 05 SAMPLES 00:02:14:00

The output.

Each reel below was produced end-to-end by the pipeline — script, footage, grade, captions. No manual editing.

script · footage · grade · captions

zero human edits

script · footage · grade · captions

zero human edits

WHAT MAKES IT DIFFERENT

Semantic, not keyword

CLIP embeds understand meaning. "melancholy evening" finds footage that feels that way — not just shots with people outside at night.

Color science, not presets

LUTs derived from footage analysis. warm-analog · cool-stark · raw. FFmpeg pipelines, not Instagram filters.

4-signal QA before edit

Sharpness, palette coherence, motion energy, brightness — every clip scored before it enters the timeline.

04 / 05 BRAINROT · SECONDARY PIPELINE 00:03:46:00

04 — BRAINROT · SIBLING PIPELINE

Brainrot.

reddit story → TTS narration → gameplay footage → karaoke captions

Takes a Reddit micro-story, generates local TTS narration, lays it over split-screen gameplay footage, and burns word-synced karaoke captions. The format that broke TikTok, automated.

brainrot — zsh

$ python brainrot.py generate --mood funny --telegram

  ↳ narration · split-screen · karaoke · $0.01

WHAT'S DIFFERENT FROM REELGEN

Brainrot is a separate pipeline optimized for the Reddit-story format. Local TTS via Kokoro-82M (Apache 2.0) — no ElevenLabs bill. Gameplay footage is provided from your own library.

Word-synced karaoke burns are the whole point. Every word highlighted on exact cue, in a font stack that doesn't look like a PowerPoint. Works offline, no API required if you skip the story-polish step.

MODES

--mood funny · touchy · self_realization

--telegram   (direct-post to Telegram channel)

--preview    (see the story, skip rendering)

brainrot · funny

brainrot · touchy

brainrot · self-realization

NOTE

Bring your own gameplay footage. The pipeline handles everything else — TTS, sync, captions. Hover to preview above (muted).

05 / 05 HONEST LIMITATIONS 00:04:52:00

Honest
limitations.

The pipeline does a lot, but it doesn't do everything. These are the real constraints, not reframed as features.

git log --oneline -- LIMITATIONS.md

a3f2e1b render time 1–3 min on CPU — no GPU acceleration (yet)

8c4d910 free LLM tier ≈ 20 videos/day — bring your own key for more

5b2a234 footage rights: your problem, not the pipeline's

7e9f123 good editing doesn't guarantee views

2d1c456 Whisper accuracy drops on noisy source audio

f4a0789 color grades tuned for warm palettes — cool tones YMMV

"Good editing doesn't guarantee views — but bad editing guarantees you won't get them."

AI video pipelines that don't look like AI.