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 30 October 2026
Hy-Nomas fits on a 4090. Its measured browser route uses eight B200s.
Project and community reports: Hy-Nomas runs on one RTX 4090, but the published browser-speed test uses eight B200s. The official 4090 preset shows that the model fits with INT8 weights and CPU offload; it does not report a viewer-speed result. In a separate Reactor run, eight B200s generated 18.53 FPS over two minutes. A relayed Chromium viewer presented 17.43 FPS, below the project’s 24-FPS video timebase. An independent ComfyUI port took 85 warm seconds to finish a 10-second 1280×704 clip on a 4090. The same port flags a noncommercial license on its PixelUMM-derived decoder weights despite Apache-2.0 wording for the upstream project. Check rights before paid use. No pool-operated long session or complete bill is available.
Choose the route by the result you need
| Route | Published evidence | What it does not show |
|---|---|---|
| Official low-memory 4090 | Tencent README: INT8 weights, FP8 decoder, CPU offload and sliding-window history. | No matched warm receiver rate or complete bill. |
| Reactor browser on eight B200s | Integration report: 2,880 frames in 120 seconds at 18.53 generated FPS; relayed Chromium presented 17.43 FPS. | Not sustained 24-FPS playback, p95 input-to-screen delay, or an accepted long session. |
| Independent 4090 ComfyUI | Port README: warm 10-second, 1280×704 completed job in 85 seconds. | Not a live viewer. Its 12-GB fit timing was simulated on a 4090. |
Rights and next test
The ComfyUI weight manifest describes its decoder as a fine-tuned PixelUMM derivative with noncommercial terms; NVIDIA’s PixelUMM card limits its own checkpoint to noncommercial research or evaluation. Tencent’s repository calls Hy-Nomas Apache-2.0. The documents need file-specific clarification before a paid deployment. This is a rights question, not a legal determination.
If you need a local visual test, try the 4090 workflow as an offline job and measure your own output quality and wall time. If you need a stream, ask for a longer remote receiver trace with control-to-visible timing and a bill. No such pool measurement is available.