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 10 November 2026 · Project and paper report
Matrix-Game 4.0 reports 112 ms from keypress to screen. It does not render any game.
Authors report: a single RTX 5090 runs the scene-specific hybrid game at 480p and 60 FPS, with 112 ms from keypress to display. A conventional low-poly game engine handles rules and controls; the video model changes how its frames look. This is not a free-form world made from a prompt or a tested overlay for another game.
Which number can you use?
| Published result | Condition | What remains open |
|---|---|---|
| 60 FPS, 112 ms | Paper's RTX 5090 game and display result at 480p; author-reported. | Independent receiver trace, long-session quality and full cost. |
| 67 FPS / 58 FPS | README's RTX 5080 / 5070 Ti with SageAttention, fixed prompts, 864×480, warm model throughput. | Video I/O and browser display are excluded; no matching input-to-screen result. |
The browser demo uses WebRTC but is a trusted single-user route, not a public cloud-gaming service. Castle, Ancient Egypt and Racing have separate released scene resources. The maintainers say these models do not yet work as a plug-and-play renderer for arbitrary Roblox games. Read their reply.
First useful test
On a supported GPU, compare one released scene with its unmodified low-poly game. Record cold readiness, browser frame rate, action-to-changed-frame p95, image stability during a longer play session and the complete bill. Keep the 5080/5070 Ti test on fixed prompts; their model-FPS figures do not establish the paper's 5090 response.
Unknown: a costed, accepted hour of gameplay, independent visual-quality ratings and a general-purpose game integration.