An infinite AI livestream from the fal camp, and the code is open source
Twitch chat steers the scene, an LTX model generates nonstop video on fal, FFmpeg pushes the stream. The whole system is on GitHub under MIT.
A striking open-source project has landed from the fal ecosystem: an AI livestream that never stops and bends to whatever viewers type. The system, called Infinite TV, is published on GitHub under the MIT license, authored by Alex, who also maintains fal's demo repositories.
The clever part is not trying to render video frame by frame, but wiring a few modules into one loop: viewer messages are collected from Twitch chat; a language model (OpenAI, or Groq for faster inference) turns them into the prompt for the next scene; the LTX video model generates a new few-second clip on fal's serverless GPU infrastructure; FFmpeg stitches the clips and pushes them to Twitch over RTMP. While the current clip plays, the next one is already being generated.
The whole equation rests on one condition: as long as generation speed stays above playback speed, the stream never breaks. Every viewer message can bend the story on the spot, which means what is on screen is not a recording but a scene that lives alongside the chat. The package even ships a dashboard for generation metrics and queue status.
It is a strong signal that the next stop for AI video is not polished clips but interactive streams. Opening the code under MIT means anyone can stand up their own "infinite television" on their own channel; all it takes is Python, FFmpeg and a few API keys.
