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 13 October 2026 · Project release
StreamRender-H3 has a browser loop. Its 218 ms figure is not player delay.
Project says: a racing game sends semantic frames to a two-step H3 renderer and receives generated frames in a browser. The game engine controls the track, cars and physics; H3 controls their appearance. The tested deployment uses eight GB200 GPUs. This is a released integration, not a pool-run session.
Three clocks, three meanings
| Reported clock | Boundary | Do not infer |
|---|---|---|
| 218 ms | Warm two-step diffusion render on eight GB200s. | Keypress-to-screen time. |
| 295.58 ms median | GPU-resident RGB input to RGB-ready output in longer validation. | Internet viewer delay. |
| About 1.03 s | Intervals between browser publications in that validation. | One-second control response or 24-fps browser feedback. |
The released asset card says the browser integration ran successfully in one same-rack eight-GPU setup. These clocks use different boundaries; they should not be added or substituted for a measured control-to-visible result. The project page gives the render result and demo.
Where this route fits
This checkpoint was trained for a third-person racing scene. The code/semantic-map design may extend to other games, but this release does not prove it. The assets total about 217 GB, and the card lists MiniMax H3 Community License terms for H3 derivatives. A consumer-GPU setup, commercial right for a particular use, and cost per playable hour are Unknown.
Next test: on an authorised deployment, log a keypress, the first changed browser frame, p50/p95 publication gaps, a restart, and the full allocation bill. That result would tell a builder more than render FPS alone. The pool has not run or paid for this route.