Shaduf.Research preview
Real-Time AI Video/StreamForce: an unpaced change sat 13 seconds ahead of display in a local test

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 15 October 2026 · Project release and paper

StreamForce: an unpaced change sat 13 seconds ahead of display in a local test

Project reports: its released browser demo accepts point pushes and wind changes while video is generated from one image. In an unpaced local test, a force change landed 213 frames ahead of the viewer, about 13 seconds at 16-fps playout. The latest paced test put the effect six frames ahead in steady state. That is a queue position, not a measured remote input-to-screen delay.

What the tests measured

ConditionPublished resultLimit
One H200, newer demoAbout 23 generated FPS at 480×832.Generation rate, not player response. Browser playout is 16 FPS.
501-frame local headless viewer, no pacingForce effect 213 frames ahead of displayed frame.About 13 seconds of queued playout, not model processing time.
Local pacing, three steady-state runsEffect six frames ahead, about 0.4 seconds; one or two startup-only stalls.The first post-launch change was 23 frames ahead in a separate run. No remote-link p95.

The demo notes explain the trade-off: holding generation near the viewer reduces stale controls, but holding too long drains the playback buffer. The June paper reported up to 16.6 FPS on one GPU; the newer repository’s H200 numbers are a different implementation and test, not a matched speedup.

Where the experience stops

The repository estimates roughly 40 GB of GPU memory for its single-GPU demo without automatic captioning; its faster path targets H100/H200. The default 501 frames are about 31 seconds at 16 FPS, but the authors report quality drift beyond roughly two to three times the 81-frame training horizon: about 10–15 seconds of playout. Their force control steers motion; it is not a physics simulator.

Try next: on the intended GPU and network, record a changed force’s input time and first visibly changed browser frame, stalls, scene quality after 10 and 20 seconds, and the complete session bill. Until that test, StreamForce is a concrete short interactive experiment, not a proven long-running world or priced service.

Search published pools, pages, reports, and evidence.