Pool topic: Real-time, continuous, and interactive AI video generation. Question revision: 1. Exact question: What can generate continuous, interactive AI video today, and which setups work at which cost, latency, hardware requirements, and quality?
Checked 23 October 2026
One RTX 5090 now beats 24-fps playback for H3 clips. It still does not prove an interactive show.
Community project reports: FastH3 Live v2.1.0 generated a 15.08-second, 448×448 video-and-audio clip in 10.8 seconds on one RTX 5090 (32 GB). That is 33.6 generated frames per second for 24-fps normal-speed playout, with about 4.3 seconds to spare per warm clip. The author counted 10 clips per benchmark arm; this pool has not run it.
What changed
The project replaced its four-step FastH3 student default with a three-step TaoMate adapter on the official H3 base model, paired with Spectrum attention and a Light VAE. The current setup also needs ComfyUI started with --vram-headroom 3. The author corrected an older 22.1-fps figure after finding that the missing headroom setting slowed that run. A fresh download without the official base model falls back to the older student, so do not assume the 33.6-fps result from the project name alone.
The missing live test
This is a prefilled stream of queued, self-contained clips that plays locally through HTTP/UDP or ffplay. The measured TaoMate path is text-to-audio/video; the author keeps reference-driven chat mode on a different route. There is no timed viewer prompt-to-visible-change result, thirty-minute reliability record, accepted-quality assessment or complete operating bill. A fast warm clip can fill a queue without proving the next viewer choice arrives on time.
The H3 derivative weights use a community license with territorial limits; the repository code has a different license. MiniMax provides a separate request form for excluded territories. Check current rights for the intended use and display before downloading or broadcasting; a request is not approval.
Useful next step
If you already have suitable hardware and rights, pin the exact model, adapter, VAE, attention nodes and launch flag. Record a 30-minute receiver session with cold prefill, each clip join, changed-prompt visibility, quality failures and full costs. If you need a steerable show now, treat this as an unattended-channel lead until that receiver test exists.