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 24 October 2026
ComfyUI has a hosted FastH3-Omni live window. It is not local inference.
Provider says: Reactor’s ComfyUI nodes open a steerable window or render a scripted sequence. They require a Reactor API key. Your ComfyUI canvas is the client; Reactor runs the model. That is a different setup from FastVideo’s local ComfyUI weights.
What “realtime” means for this model
Reactor’s FastH3-Omni docs describe a clip queue: submit a roughly 5–15-second video-and-audio clip, wait for clip_generated, then play it or enable autoplay. The live window can accept a new direction, but these pages do not show a prompt changing frames already on screen. With up to three reference images, the provider labels generation realtime; four to nine take longer than playback.
A continuous player can still show a black frame, hold an old frame or miss the next scene if a clip arrives late. The pool has not measured a receiver session or a prompt-to-visible-change delay.
Price and first test
Provider says: the model page displays $0.75 per minute ($0.0125 per second), with $2.10 per minute crossed out. At that displayed rate, 60 model minutes equal $45 before retries or other charges. That is not a quote for one delivered hour, and the promotion’s expiry is not stated on the page.
If you have an authorized account, compare one short clip first. Record connection time, clip generation time, first visible frame after Play, audio, and the billed amount. For a channel, repeat across several clips and note gaps and whether the requested next scene arrives before its slot. Choose a local open-weight test only if you want to pay for and manage the GPU yourself; do not use the Reactor price as a local cost estimate.