Narrated text-to-video on a CPU, with no GPU and no API key

Most "text to video" tools either call a paid cloud API or need a GPU. Agent sandboxes and CI runners usually have neither. makevideo renders on a plain CPU:

Voice: Kokoro-82M (Apache-2.0) running locally through ONNX Runtime — or a Piper voice with --fast that speaks ~10x faster than real time on two CPU threads.
Frames: drawn with Skia (@napi-rs/canvas) — no headless browser.
Encoding: H.264 via a bundled ffmpeg (ffmpeg-static).

npm i github:random00000000/makevideo
npx makevideo setup          # pre-download voice models once (CI / cloud images)
npx makevideo video.md       # or video.json
npx makevideo video.md --fast

How fast is it?

Measured on a 6-core laptop CPU (i7-10750H), no GPU, for the same 74-second, 8-scene video:

	time
first render, default voice	~81 s
first render, `--fast` voice	~36 s
re-render with cached speech	~28 s

Why it is fast on a CPU

Frame holding. Slide videos are mostly still frames. makevideo knows each scene type's animation timing, draws only frames that can change (about 30%), and lets ffmpeg repeat held frames after colour conversion, where repeats are almost free.
Render processes. Scenes draw and encode in a pool of processes, so drawing uses many cores.
Tuned speech. ONNX threads are pinned per process; processes that each grab every core fight and get slower.
Caching. Every sentence's audio is cached by voice + text, so re-renders after a layout fix skip speech entirely.

From code

import { video } from 'makevideo';

const r = await video({
  title: 'Patch notes',
  scenes: [
    { type: 'title', title: 'Patch 1.2', say: 'Here is what changed in patch one point two.' },
    { type: 'chart', title: 'Load time', unit: ' s', bars: { Before: 12, Now: 3 }, say: 'Load time. Twelve seconds before. Three now.' },
  ],
});
console.log(r.output, r.duration, r.seconds);

← makevideo home · GitHub