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 7 October 2026 · H3 speed claim check
Veda cut this H3 render to 130 seconds. The video lasts five.
Project says: Veda-on-ComfyUI makes H3 attention faster, but its 7.1× figure applies to attention time per step, not the complete clip. The project's complete-job test improved 2.63× from the default ComfyUI path.
The matched test
On an RTX 5070 12 GB under Windows 11, the project timed a 124-frame, 1344×768 H3 T2VA job with an eight-step Turbo LoRA. That is about 5.17 seconds of output at 24 fps. Its measurement log reports 342.1 seconds with default dense attention, 230.9 seconds with SageAttention, and 130.1 seconds with Veda. Dividing 342.1 by 130.1 gives 2.63. The reported 7.1× compares 31.1 and 4.41 seconds of attention per step. Inspect the project log.
The Veda job still took about 25 wall seconds for each output second. That is useful for offline iteration, not playback-rate generation or an interactive stream.
What to test before switching
A 3060 user reported an eight-minute run for a different eight-second, 0.5 MP, six-step workflow, with poor audio and no apparent improvement over their usual method. That is not a matched comparison against the 5070 result. Pin your ComfyUI version, model, LoRA, seed, resolution and frames; compare complete playable clips and accepted audio and picture on your own hardware. Do not infer a 5090 speed or a delivered-stream cost from this one test.