<3s
to generate ~5s of video
Faster than a short scene can finish playing.
Infinite TV
Never go off air.
While viewers watch this scene, the next one is already being written and rendered.
The insight
Infinite live is not frame-by-frame realtime.
It is a pipeline that stays ahead of the playhead. As long as generation finishes before the current clip ends, the stream never drops.
Headroom becomes continuity. Continuity becomes infinite TV.
The loop
Five modules. One closed circuit.
Infinite AI live is a relay: listen, rewrite, render, stream, and stay one scene ahead.
Chat in
Twitch messages arrive from the live audience.
Rewrite
An LLM turns chatter into the next-scene prompt.
Generate
A fal video model renders a few seconds of footage.
Mux & push
FFmpeg stitches the clip and pushes RTMP to Twitch.
Prefetch
The next clip generates while the current one plays.
The ring never breaks — as long as gen > play
Speed
Generation that outruns the playhead.
MiniMax H3 Max on fal turns AIGC from “make then publish” into something that can feed a live RTMP stream.
~9s
to output ~15s of video
Enough headroom to keep the buffer warm.
10×
throughput vs stock H3
Post-training and inference work on fal.
When generation finishes before playback ends, the stream can run forever.
Audience as director
Every message can rewrite the next scene.
Chat does not decorate the stream — it steers it. Viewers become co-authors of whatever happens after the current clip.
Twitch chat→ next prompt
- novamake it rain neon in the alley
Scene prompt
Open source
Build your own endless room.
The infinite-tv pipeline is public: chat in, prompts out, video on fal, FFmpeg to RTMP. Fork it, point it at a stream key, and keep the lights on.
LLM promptsfal video modelsFFmpegRTMPTwitch
Open the repository