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 29 October 2026
SGF+ generated 24-hour videos. Its live browser route is a different test.
Authors report: four 24-hour SGF+ rollouts from a model trained on five-second windows. That is a long-generation result, not a viewer session. A separate Reactor browser integration reports about 19 generated FPS on one B200 while playing at 16 FPS. An independent ComfyUI pack ran a 20-second 5090 file job in 56.6 warm seconds. Choose the route for the job, not the largest duration in the headline.
Three results, three decisions
| Route | Evidence | Decision |
|---|---|---|
| Research rollout | Paper: four 24-hour examples; main quality evaluation at 60 and 240 seconds. | Study long-horizon generation, not a proved 24-hour broadcast. |
| Reactor browser | Project reports one B200, 832×480, about 19 generated FPS for 16-FPS playback. Prompt edits affect a later chunk, after buffered frames. | Try only if you need an interactive browser and can test response and cost. The default horizon is about four minutes before a new start. |
| ComfyUI file | Integration author reports 321 frames at 832×480/16 FPS in 56.6 model-resident seconds on one RTX 5090. | Use for an offline quality check, not a live feed. |
What to check before a show
The B200 guide reports a 30-chunk browser test, not an accepted hour. Unknown: first-frame delay, p95 prompt-to-visible time, longer-session quality, recovery and the full bill. A useful next test would time a short controlled session, inspect visible quality after each prompt change and keep the invoice. The ComfyUI job and paper timer use different hardware and output contracts; do not rank them by FPS.