The studio makes presentation videos from a simple text brief. Since September it has two hands, and they do not do the same job.
- The hand that speaks — HyperFrames: HTML → MP4, deterministic frame by frame. Everything that must be right comes from here: a title, a number, a date, the tempo.
- The hand that acts — MiniMax H3, served by ComfyUI on the cluster: image → video with speech. Everything that must be alive comes from here: a face, a light, a mouth articulating syllables.
- The result: a self-hosted character introducing himself, subtitled with exactly what he says — because the letters come from the other hand.
- Visual identity: the heart with a wink logo 💚😉, and the team's face, drawn here.
We publish our limits too: the animation model repaints
letters (its character holds a sign reading "ASTER" and pronounces it "Astaire") and will not
hold a static shot. So every image is re-read before it is shown,
and what fails is kept in a drawer called rats/, with its reason written
out — never erased.
The production pipeline relies on a community open-source tool — heygen-com/hyperframes (video rendering from HTML) — which we thank and cite as it deserves. GreenCluster is built to be shared and reproducible : everything is in the git, for anyone.