ComfyUI Spectrum Node Cuts MiniMax H3 Generation Time by ~30%

ComfyUI Wikinews

xmarre's ComfyUI-Spectrum-MiniMax-H3 node uses spectral feature forecasting to skip expensive transformer evaluations, delivering ~30% faster MiniMax H3 generation in ComfyUI.

xmarre's ComfyUI-Spectrum-MiniMax-H3 brings spectral feature forecasting to ComfyUI's native MiniMax H3 audio-video model. By forecasting post-transformer hidden features and skipping selected transformer evaluations, it delivers roughly 30% faster generation without loss of quality in community tests, and has already racked up over 200 GitHub stars within days of release.

The node is an independent port of the Spectrum acceleration method (Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration) for the H3 omni-modal model that MiniMax open-sourced on August 3, 2026 with native ComfyUI support (read the open-weights story).

How It Works

During sampling, the node fits a Chebyshev ridge regression model to the actual post-transformer hidden features of H3, then uses it to forecast those features on selected future solver steps. On forecast steps, the expensive transformer blocks, RoPE construction, conditioning projections, and reference embedding are all skipped — only the native output heads, video/audio reconstruction, and sigma mapping still execute.

The method was designed for diffusion sampling in general:

Spectrum paper samples: 3.5x speedup with only 14 network evaluations

Spectrum paper samples (top: text-to-video on HunyuanVideo, bottom: text-to-image on FLUX.1), generated with only 14 network evaluations for a 3.5× speedup.

Speedups

On a 20-step run with the conservative preset, the sampler-aware scheduler skips 6-7 of 20 H3 transformer evaluations:

SamplerActual stepsForecasted stepsTransformer-step reduction
Euler13735%
RES multistep / CFG++14630%

Community reports match those numbers: BecauseReasons shared the node with a "~30% faster without quality loss, unlike cache" verdict, and nghtdrp confirmed similar results. An AMD ROCm user reported a 20-step RES multistep run (864×480, 107 frames) dropping from 212.7s to 161.0s — about 1.32× throughput on a Radeon AI PRO R9700 with ComfyUI 0.30.0.

Quality Caveats

Spectrum is an approximate accelerator, not a lossless path. The README is explicit that forecasted steps change the denoising trajectory:

  • During fast or brief actions, motion, pose, timing, or gaze can diverge from a fully native run
  • Rapidly moving or briefly visible details (eyes, fingers, fingernails) can occasionally become malformed or unstable
  • Outputs can differ even with identical prompt, seed, model, and workflow

For quality-critical work, compare the same prompt and seed with the node enabled and disabled.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3.git

Restart ComfyUI. The node appears under sampling/spectrum as Spectrum Apply MiniMax H3, with no third-party Python dependencies.

Requirements: ComfyUI with native MiniMax H3 support (0.30.0 or later, native H3 merged August 3). The node accepts and returns MODEL; recommended placement is after the H3 Sigma Shift and before the guider and sampler. Euler, RES multistep, and RES multistep CFG++ are supported; ancestral samplers and multi-GPU parallel sampling automatically fall back to the native H3 path.

Comments

Sign in with GitHub to join the discussion.

Loading comments…
ComfyUI Spectrum Node Cuts MiniMax H3 Generation Time by ~30% | ComfyUI Wiki