SPEED ComfyUI nodes generalize to any sampler: low-res early diffusion, up to 1.58x faster
antipode_insights · reddit · 2026-11-03
The author generalized their MiniMax-H3 SPEED nodes into universal ComfyUI-SPEED, a 1:1 drop-in replacement for KSampler/SamplerCustomAdvanced that works with image, edit, and video models (thoroughly tested on H3, MOSS, Krea 2; compatible with FIBO 2 and Flux 3).
How it works: runs the early, noisy diffusion stages at low resolution, then steps back up to full resolution as detail starts to matter.
What's new
- Image model support (MOSS Image, Krea 2, other flow/CONST models) and most samplers
- Ready-made workflows for H3, MOSS, Krea 2 with harvested sigma values — load and queue
- Tooltips on every input; old node IDs kept as deprecated for backward compatibility
- Transition noise now drawn on CPU: same seed gives same image on any GPU
Benchmarks (same seed/prompt):
- MiniMax-H3 5s 0.5MP clip, L40S: 288s → 182s (1.58×)
- MOSS Image 1344×768, RTX 5070 Laptop: 22s → 14s Manual (1.54×)
- Krea 2 1216×832: 283s → 183s Manual (1.55×)
Run Sigma Harvest once per model/sampler/schedule. Repo: StanLukuvka/ComfyUI-SPEED.
Related event: ComfyUI-SPEED: Training-Free Sampling Speedup for Diffusion Pipelines(2 posts)→
More from Multimodal
- AssemblyAI throws haunted TTS party to show off new voice model — AssemblyAI · 2026-11-04
- Hedra speeds up Krea 3 diffusion transformer 6.9x on AMD Instinct MI355X — henloitsjoyce · 2026-11-04
- PrunaAI Open-Sources P-H3-Image: 7x Smaller, 5x Faster, 8-Step Images — sdiazlor · 2026-11-04
- AssemblyAI launches first TTS model on LiveKit: phone-call-trained, 90ms first audio, $15/1M chars — AssemblyAI · 2026-11-04
- Fully AI-edited 'After Hours' short built with open-source engine OpenKatana — rk0neki · 2026-11-04
- From one 82-second video to a 3D Gaussian splat in ~20 minutes — janusch_patas · 2026-11-04