Endless AI TV channel on one RTX 5090: MiniMax H3 generates faster than it plays
spartong945 · reddit · 2026-09-02
A Redditor built a never-repeating, locally streamed AI TV channel on a single RTX 5090 using open-weights video model MiniMax H3 via ComfyUI — picture and synchronized audio generated from text, played in VLC from a local URL, no cloud.
Key details:
- Uses the 4-step FastH3 distillation; each clip is 362 frames (15s). Generation must outrun playback forever: 20.1s of video per 19.2s of GPU time. Honest caveat: played at 18fps vs native 24fps, so motion runs at 75% speed.
- Weights quantized from 66GB to 21GB INT8; with sage attention and an INT8 VAE a 15s clip took 26.5s. A 40-line node-level profiler using ComfyUI websocket events revealed SaveVideo (PyAV per-frame Python loop) ate 3.78s — a fifth of each run — and was replaced with ffmpeg.
- Content: 321 hand-written scenes, 503 characters, randomized ensembles; weights and code are on Hugging Face (jacokon/fasth3-live).
Related event: MiniMax H3 Powers Infinite AI TV Streams Faster Than Playback(3 posts)→
More from Infra
- Auction-based LLM routing paper finds fixed token pricing wastes 10%-71% in costs — _onionesque · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
- SGLang Summit sponsors jump from 11 to 36 in one week — ying11231 · 2026-10-18
- Oracle CDS hits record 300bps, called first AI trade domino after Barclays sell downgrade — arthurcolle · 2026-10-18
- Agent Coding Workflow Would Cost ~$40k/Month on API Pricing, Dev Reveals — ChrisRackauckas · 2026-10-18
- PS5LM: llama.cpp port turns a jailbroken PS5 into a home LLM server (500+ GitHub stars) — wateriscoding · 2026-10-18