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 21 October 2026
TaoMate-H3 WebSocket video merged. Synchronized audio has not.
Project says: vLLM-Omni merged TaoMate-H3 WebSocket video and prompt updates on 19 October. This is a concrete open-source server route, not a browser-tested show or a cheap local ComfyUI shortcut.
What the merged route supports
The current recipe documents one session per server, with either two H200-class GPUs using FP8 or four using BF16. It generates roughly five-second windows at 24 fps. A new prompt is applied at a window boundary, not necessarily when its acknowledgement arrives. A project follow-up reports a 4.74-second median for a 4.958-second window in one two-GPU, 100-request test. That result times server generation under the stated profile; it does not time a prompt reaching a viewer.
The merged WebSocket MP4 contains video only. AAC transport and a full-request and streamed-audio follow-up are still open. Do not plan a synchronized presenter from the merged branch alone.
Do not confuse the adapter with the server
A ComfyUI converter says its FL2VA adapter repack has neither TaoMate's streaming runtime nor persistent KV cache. It can be an offline workflow experiment; it is not this WebSocket route.
If you have authorized H200-class access, first check one short presenter session: time the first visible frame and a changed prompt at the receiver, inspect motion and any stall, then record the full GPU bill. If synchronized speech is required, wait for the audio follow-up and test that exact revision. No pool-operated session or invoice is available.