FastH3 + USDU doubles video upscale speed on RTX 3090: 48m vs 96m, full workflow released
inazma44 · reddit · 2026-09-07
After days of testing with an AI agent, the author combined FastH3 (MiniMax H3) with Ultimate SD Upscale for restoring low-quality AI generations, and published a reproducible workflow.
Benchmarks (15s source, 2752×1536, 2 steps, denoise 0.20, RTX 3090 24GB):
- H3 + 4-step Turbo LoRA / USDU: 96m 37s
- FastH3 / USDU: 48m 40s (2× faster)
- FastH3 / MMH3 single-node upscale: 47m 36s
Verdict: FastH3 + USDU wins — the single-node path saves 1 minute but introduces more distortion; both final paths finish with pixel-domain ESRGAN upscaling.
Caveats: single successful runs with different cache states, not a strict model-only benchmark; FastH3 USDU hit two OOMs requiring a memory-estimation override. A 5-second clip also ran on an RTX 4070 Ti SUPER (16GB VRAM).
The full workflow and Windows installer (English instructions, 41GB model links) are on GitHub; read the installer notes — it updates Kitchen and applies fingerprint-guarded H3 core patches.
More from Multimodal
- Building an entire AI-generated cinematic world with CineMap and TapNow_AI — egeberkina · 2026-10-26
- Claude Code can now edit videos: a 10-prompt workflow from raw footage to final cut — TawohAwa · 2026-10-26
- One-sentence 3D worlds turn cinematic with CineMap and TapNow — gizakdag · 2026-10-26
- No Camera Needed: Creator Builds an Entire World in CineMap and Directs It Into Cinema — gizakdag · 2026-10-26
- Kling vs. Seedance: the Chinese AI video model rivalry — DavidmComfort · 2026-10-26
- Bokeh Cosmic Nebula prompt template: out-of-focus orbs mimic interstellar clouds — LudovicCreator · 2026-10-26