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 12 October 2026 · Existing route, not a new release
TeleFuser made 60 seconds of media. The client session took 79 seconds.
Project report: TeleFuser has a real LiveKit/WebRTC browser route for LingBot-World v2 on four H100 80 GB GPUs. Its one-minute replay generated 59.75 seconds of media, but the measured session lasted 79.2523 seconds. The first client frame arrived at 5.9822 seconds. That is a more useful live-video check than model FPS alone. It is not a pool-operated test or proof of a continuous hour.
Two tests, two clocks
| Project test | What it measured | Limit |
|---|---|---|
| 77-frame gate, commit 540b579 | 832×480, 16-FPS target; 17.1399 steady target compute FPS on four H100s; first generated frame at 3.2182 seconds. | Excludes model loading, LiveKit pacing/encoding, network and client rendering. |
| Later one-minute LiveKit replay | 957 generated frames, 59.75 seconds media; 17.5824 steady target compute FPS; 5.9822 seconds to first client frame; 15.2313 client callback FPS after first frame; 79.2523 seconds session runtime. | One successful project-maintained run. Three extra decoder callbacks were startup frames, not generated model frames. |
The tests use different workloads and code revisions. Do not read the first row as the receiver result for the second. See the project benchmark and conditions.
What to test before deploying
The repository example provides image input and keyboard controls through a browser, but caps the example session at two minutes. A one-minute successful replay does not establish prompt- or keypress-to-visible delay, freeze behavior, accepted visual quality, a one-hour handover, multi-viewer load or delivered-hour cost. Test one fixed scene and control script at the receiver. Log first frame, each control event, frame cadence, a failure/reconnect and the bill. Compare on the same task; TeleFuser's separate H3 service route is not this LingBot stream.