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 16 October 2026 · ComfyStreamerH3 release
One RTX 5090 beats playback speed. The browser still waits for two clips.
Project reports: a warm RTX 5090 generated two 15.08-second H3 clips at 448×256 in a median 23.51 seconds across 11 jobs. That is 30.16 seconds of video in 23.51 seconds, or 1.28× playback speed. It is not a measured prompt-to-screen time or proof that the one-job-at-a-time HLS loop can keep up.
The two clocks
| Clock | What is published | What remains unknown |
|---|---|---|
| Warm generation | Two 15.08-second clips in a 23.51-second median paired job on one RTX 5090, four-step VSA. | Per-clip delivery times and accepted-quality rate. |
| Browser playout | The released HLS demo waits for two completed MP4s, then segments them locally. A new prompt waits until the current clip finishes. | Time from prompt to first changed frame at a viewer, startup, stalls and long-session behavior. |
The README gives the warm throughput figure. The viewer source shows the serial workflow: wait for a completed remote MP4, convert it to HLS segments, and start playback only after two clips exist. Its player requests a 30-second live-sync target, but no receiver delay is measured. The author's engineering note says the low-resolution clips have visible artifacts and need better detail and consistency.
Which setup costs what?
The 5090 result uses a specific pruned FastH3 V2 workflow; do not transfer it to full H3 or a higher resolution. The README's managed-deployment example selects an RTX PRO 6000 rather than a 5090. Comfy API lists that GPU at $4.54 per hour as checked on 16 October, plus a paid plan and model storage. That rate is not the cost of an accepted clip or delivered hour. The author's cheaper 5090-rental example is a different access route, not a Comfy API price.
Try next: record the first playable browser frame, a mid-clip prompt change through the first visibly changed frame, one late or failed render, output quality, and the complete bill on one pinned setup. Until then, this is a promising low-resolution clip-throughput result with a released HLS prototype, not a proven low-latency interactive stream.