SPEED ComfyUI node speeds up MiniMax-H3 up to 70% without retraining
antipode_insights · reddit · 2026-08-30
A dev released ComfyUI-MiniMax-H3-SPEED, a ComfyUI extension that accelerates MiniMax-H3 image generation with no retraining. The idea: research shows early diffusion stages carry no fine detail, so running them at lower resolution yields big speedups with little quality impact. Conservative settings give 20% speedup with no quality loss; aggressive settings reach up to 70% but produce basically unusable outputs (still useful for seed inspection and rough prompt drafting).
Three nodes ship: Sampler (Automatic) picks a 2/3/4-stage schedule with a baked 1% delta for default H3; Sampler (Manual Step-Through) lets you set up to four (goal, resolution) pairs, e.g. to copy a paper's schedule; Sigma Harvest runs a native Euler pass to measure the noise spectrum of your specific model/LoRA combo and outputs A / β / Δ — re-run it whenever you change model or LoRA, as defaults only fit base H3.
Roughly compatible with anything that doesn't touch the sampler directly, though only base H3 and Turbo LoRAs were tested; Euler-only sampling as in the paper.
Related event: SPEED ComfyUI Node Speeds Up MiniMax-H3 by 20% Losslessly(2 posts)→
More from Infra
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- Gemma 4 31B QAT on one 24GB card: 46 tok/s plain, ~80 with draft model on RTX 4090 — Subject_Poet_3861 · 2026-10-16