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 19 October 2026 · H3 serving measurements
The reported H3 slowdown disappeared. Do not price a route from that CI run.
A vLLM-Omni nightly test showed a four-H100 H3 text-to-video job taking 142.84 seconds against a 38.33-second baseline. The issue closed after the metric returned to normal the next day. The author suspected machine contention, not a confirmed software regression.
Why the comparison changed
The failing test used 1344×768 output, 209 frames, eight steps and four H100 GPUs. A contributor ran the two commits with the same request shape on four H200 GPUs. That rerun did not reproduce the slowdown: its text-to-video times were 49.18 and 46.12 seconds. H200 is not H100, so this does not prove what happened on the original host. The failing log showed one card running unusually hot and drawing more power; complete telemetry for the failure interval was not published.
Practical rule: before using a single benchmark to choose or rent hardware, check the same model, task, commit, GPU type, host condition, warmup and complete-output timing. If a large change vanishes on the next run, keep it marked unresolved until a controlled rerun explains it.
A new API optimization is not a new H3 speed
An open 19 October pull request proposes sending encoded frame counts from the H3 worker to the video API. That could avoid reopening a completed MP4 just to count frames. The contributor reported no controlled end-to-end speedup. A reviewer noted that the split-decoder route would still use the old count path and asked for a before/after measurement. This PR does not show faster generation, earlier first frames or shorter viewer delay.
The October runtime roadmap also proposes backpressure to stop generation running too far ahead of a client. Its authors state that backpressure cannot fix playback underruns when inference is slower than playout. Both improvements may matter for serving, but neither supplies a priced, sustained H3 viewer session.