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 28 October 2026
Micro-World 2 reaches 42.49 decoded FPS on one MI350X. What reaches the viewer is still unknown.
AMD reports: its new causal world model generates 352×640 video from a starting frame, prompt and keyboard/mouse actions. On one Instinct MI350X, the optional lightweight decoder reaches 42.49 decoded FPS; the original Wan VAE reaches 32.12. Both exceed 30-fps playback in AMD's test, but neither figure times a browser or a bill. See AMD's technical report.
The decoder changes the answer
The lightweight decoder cuts AMD's per-chunk decode time from 124.19 to 32.05 milliseconds. AMD also reports a modest loss in image quality and pose accuracy. Its main WBench comparison uses the original decoder, not the 42.49-FPS mode. Keep speed and quality attached to the same configuration.
The checkpoint card lists separate base Wan2.2 weights and labels the model OpenRAIL. The repository says version 2.0 is still under development. The reported result is on MI350X, not an RTX 5090 or a consumer Radeon. AMD describes generation for up to one minute; no receiver-timed five-minute session, p95 keypress-to-screen delay, or complete delivered-hour cost is published in these sources.
A useful first test
If you have authorized MI350X access, compare both decoders on the same input and action trace. Record cold start, delivered FPS, keypress-to-visible p50/p95, visible defects, interruption recovery and the complete bill. Stop if the original mode misses your playback rate or the faster mode fails your visual bar. Do not buy a GPU from a decoded-FPS headline alone.
A separate StreamEval benchmark scores live world-generation quality, instruction following and continuity separately. It does not yet score Micro-World 2 and does not measure whether viewers want to keep watching. Use those categories as a test checklist, not a cross-benchmark ranking.