← sources

SPEEDing up MiniMax-H3 without retraining - V2, now with more samplers and considerably less jank

source: reddit (reddit:StableDiffusion)published 2026-09-20 01:34ingested 2026-09-21 02:37

retained text

A couple weeks ago I posted my SPEED implementation for MiniMax-H3, which lets you run the early parts of diffusion at lower resolution and then progressively move back to full resolution.

I have now done the very sensible thing and rewritten large chunks of it.

Repo: https://github.com/StanLukuvka/ComfyUI-MiniMax-H3-SPEED

Changes:

1. Fixed a little, tiny... massive fundamental issue with the code that caused the node to degrade much faster than it should have.
Oops.

1. Added sampler support. It now supports: Euler, Heun, DPM2, Exp Heun 2 X0, RES Multistep

1. A small ocean of code cleanup, bug fixes, validation fixes, test cleanup, and comment cleanup.

It should now run much more consistently, contain considerably less jank, and generally produce less weird artifacting/blurring/whatever

There are also a bunch of smaller changes around I2V handling, Sigma Harvest, sampler state, stage planning, and failure cleanup.

document fetched 2026-09-21 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1wl3mux/speeding_up_minimaxh3_without_retraining_v2_now