fifi-video

Video tools that are just code. Every frame is HTML / Canvas / JS, rendered frame by frame in Chrome. No generative video models.

English · 中文 · For AI agents

Full promo:

What's open

	Module	Status
	motion 29 animation components + render pipeline	✅ v0.1
	Rough cut for talking-head videos	soon
	BGM & sound effects	soon
	Cover prompts	soon
	Titles	soon
	prompts/ Full one-prompt videos	✅

Why these components

• Hand-picked with human feedback, RLHF-style. I built each one through rounds of render → frame-by-frame review → rating → fix. Only the ones rated A or B ship; the rest were cut. That's why they don't look AI-made.
• Reusable. A component is only a frame. Words, images, colors and timing are all parameters, so the same component works for any topic.
• Any AI can use them. The design lives in the components. The agent only fills in words and times, so you don't need the newest model. Any coding agent that can read files and run shell commands will do. Tested: given only this repo, Claude Sonnet 5.5 and Claude Haiku 5.5 each made a 15-second promo that passed every build check.

Quick start

Point your coding agent (Claude Code, Codex, Cursor…) at this repo and say:

Use fifi-video/motion to make a 10-second 1920×1080 intro that says "Hello, world".

Or by hand:

cd motion && npm install
bash scripts/new_project.sh ../my-video
bash scripts/build.sh ../my-video      # frames → sound → my-video.mp4

Needs Node 18+, Google Chrome, Python 3 (numpy), ffmpeg.

Contributing

Bug reports, fixes and new components are welcome. See CONTRIBUTING.md.

License

Code: MIT. Bundled fonts keep their own OFL / Apache licenses (motion/components/explainer/xm-assets/fonts/).

Made by 小火火Fifi.