Repo

ComfyUI-Spectrum-MiniMax-H3

Skips selected MiniMax H3 transformer passes by predicting final-block hidden features, while keeping native heads, reconstruction, and fallbacks active.

GitHub ↗Aug 3, 2026Paper ↗

  • multimodal
  • comfyui
  • diffusion-sampling
  • feature-forecasting
Comparative grid showing method vs baselines with speedups and quality across prompts
Comparative grid showing method vs baselines with speedups and quality across prompts

What it is

ComfyUI node that skips selected MiniMax H3 transformer evaluations by forecasting hidden features during sampling. Trades exact trajectory fidelity for lower runtime. Accelerates native MiniMax H3 sampling in ComfyUI by forecasting post-transformer hidden features on selected steps, then running native output heads and reconstruction every step with safe fallbacks.

Key results

  • 20-step Euler schedules 13 actual and 7 forecasted H3 evaluations
  • 20-step RES/CFG++ schedules 14 actual and 6 forecasted evaluations
  • Reported RES run dropped 212.73s to 160.97s on Radeon AI PRO R9700

Shows native audio-video diffusion models can skip some transformer evaluations by forecasting hidden states, with useful speedups but non-identical outputs.

One useful ML email a week.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.