Perception · 2026-09-21 · 1:48

You do not need a model to make the video; you need a renderer

Name: You do not need a model to make the video; you need a renderer
Uploaded: 2026-09-21
Duration: 1 min 48 s
Description: Pair a generation engine, self-hosted or pay-per-use, with a deterministic assembler you run yourself, and the pipeline is yours rather than a vendor's. Determinism is a quiet island in the stochastic flood. Build on it.

HeyGen has open-sourced HyperFrames (Apache 2.0, 52,000 stars at the time of writing): a framework that turns plain HTML, CSS and animations into deterministic MP4, same input, same frames, rendered on your own machine with no per-render fee. Agents write the HTML; FFmpeg does the rest.

In favour

It names what the "AI video" conversation hides: most of a finished video is deterministic assembly (captions, overlays, transitions, kinetic text, branding), not generative magic. The generative model becomes a swappable input.
We have used it ourselves for the book videocast: reproducible, ownable, byte-identical on rerun.

Worth watching

HeyGen gave away the deterministic engine and kept the avatar model proprietary. Give away the plumbing, keep the moat: the shape of the current AI economy in one release.
Stars are not maturity; a young framework changes fast and your pipeline changes with it.

Our takePair a generation engine, self-hosted or pay-per-use, with a deterministic assembler you run yourself, and the pipeline is yours rather than a vendor's. Determinism is a quiet island in the stochastic flood. Build on it.

Source: Η HeyGen ανοίγει τον κώδικα του HyperFrames (GitHub, Apache 2.0)

← All Focus posts

AI-narrated · AI-generated content

Transcript

What does it take to make a video with AI? You do not need a model to make the video. You need a renderer. A video company has open-sourced its rendering framework. Apache 2.0, and 52,000 stars at the time of writing. It turns plain HTML, CSS and animations into deterministic MP4. Same input, same frames. Rendered on your own machine, with no per-render fee. Agents write the HTML; an encoder does the rest. It names what the AI video conversation hides. Most of a finished video is deterministic assembly. Captions, overlays, transitions, kinetic text, branding. Not generative magic. The generative model becomes a swappable input. And we have used it ourselves, for the book videocast. Reproducible. Ownable. Byte-identical on rerun. Worth watching. The company gave away the deterministic engine and kept its avatar model proprietary. Give away the plumbing, keep the moat. The shape of the current AI economy, in one release. And stars are not maturity. A young framework changes fast, and your pipeline changes with it. So pair a generation engine, self-hosted or pay-per-use, with a deterministic assembler you run yourself. Then the pipeline is yours, rather than a vendor's. Determinism is a quiet island in the stochastic flood. Build on it.