ComfyUI-Spectrum-MiniMax-H3

ComfyUI-Spectrum-MiniMax-H3 is a custom node for ComfyUI that applies Spectrum, a training-free spectral feature forecasting method, to ComfyUI's native MiniMax H3 audio-video model. Spectrum comes from the paper "Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration" (Han, Shi, Li, Ye, Guo, Ermon). This repository is a standalone adaptation of that method for H3's packed audio/video representation.

Denoiser features are treated as functions of diffusion time and approximated with Chebyshev polynomial bases, with coefficients fitted online by ridge regression. On forecast steps the node predicts the final-block hidden state from earlier actual steps and skips the H3 transformer blocks, then continues through the native output and sampler path. It supports samplers including Euler, ER-SDE, RES, SEEDS and SA-Solver, offers CPU or VRAM history storage, and falls back to native behaviour when it cannot verify compatibility. It also lists compatibility paths with RefDelta Solver, Untwist and H3 Continuum.

Installation is by cloning the repository into ComfyUI's custom_nodes folder and inserting the Spectrum Apply node before the guider and scheduler. Because forecast steps change the denoising trajectory, output can differ from native H3 even with the same seed and workflow. Creator reports mention a version with lower quality loss, working with res multistep but not with the ERSD sampler. The README states its validation covers native CPU/source equivalence for some changes, not all GPU video output.