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 18 October 2026 · proposed Self-Forcing server
One A800 streamed its first nine frames in 1.8 seconds. It could not keep up with playback.
Contributor reports: an open vLLM-Omni pull request adds a stepwise WebSocket client for Self-Forcing, with prompt changes during a stream. On one A800 80 GB, an 81-frame, 832×480 run at 16 fps took 10.57 seconds to send about 5.06 seconds of video. This is incremental output, not an unbuffered 16-fps feed.
What the run actually times
| Reported local test | Time | Why it matters |
|---|---|---|
| First nine frames | 1.80 s | A first chunk exists; this is not a remote viewer's first frame. |
| Later 12-frame chunks | 1.16–1.74 s each | Each chunk plays for 0.75 s at 16 fps, so an empty buffer would fall behind. |
| Whole 81-frame WebSocket stream | 10.57 s for 5.06 s nominal video | About 0.48× playback rate on this eager A800 setup. |
The linked model issue describes four-step Wan2.1-T2V-1.3B and a 21-latent-frame trained window. The PR remains open. These are contributor measurements, not a pool test or a merged product benchmark.
Prompt control is not viewer response
The proposed client accepts a new prompt between generation blocks. In the contributor's test, an update sent at stream start affected chunk 1, not chunk 0; earlier cached frames still influenced the scene. No remote prompt-to-visible delay, long-session quality, multi-user load or delivered-hour cost was measured. The current path is text-to-video and does not include the camera controls of minWM; that integration is listed as later work.
Smallest useful next test: if you already have authorized GPU access and can use the open branch, cap a run at 81 frames. Timestamp first chunk, each later chunk, a prompt update and the first changed frame on a separate viewer. Stop treating it as an unbuffered 16-fps route if chunks continue to take longer than their 0.75-second playout. If you need a stable, merged route or must rent hardware first, wait for a release and a measured receiver session before budgeting a live show.