All About AI combines a fast MiniMax H3 variant, a language model, a clip queue and Twitch streaming into a pipeline for continuously generated videoContinuous video generation uses an AI pipeline to produce an ongoing sequence of clips instead of one isolated video.. At 480p on two Nvidia B200 GPUs, the demonstration produces a 15-second clip in roughly 10 to 13 seconds, allowing the system to prepare the next segment before the current one finishes.
When no viewer intervenes, the language model continues the story automatically and passes a new scene prompt to the video modelA video generation model is an AI model that creates or transforms sequences of visual frames from prompts or other conditioning inputs.. A chat command receives priority over the automatic queue, so viewers can redirect the plot while a small memory of previous scenes helps preserve some continuity.
The test exposes important practical limits. The dialogue and visuals remain inconsistent, the 480p setup costs about $14 per hourInference cost is the expense of running a trained AI model to produce outputs for real requests or workloads., and the creator estimates that a 720p version would need eight B200 GPUs and cost roughly $100 or more per hour. A continuous public stream would therefore need a viable funding model and would still face platform-policy questions.
The experiment shows that real-time generative videoReal-time video generation creates AI video quickly enough for new footage to be produced as an experience is being watched or directed. is becoming technically possible before it is consistently coherent or economical. Faster inference, lower compute costs and stronger image-to-video continuity would make interactive, continuously generated channelsInteractive generative video creates a visual sequence that changes in response to user controls or events while it is being generated. much more practical.
Watch on YouTube
