xmarre/ComfyUI-Spectrum-MiniMax-H3

Training-free Spectrum acceleration for ComfyUI’s native MiniMax H3 audio-video model. Uses Chebyshev ridge feature forecasting to skip selected H3 transformer evaluations, with adaptive scheduling, sampler-aware support for Euler, ER-SDE, RES, SEEDS and SA-Solver, CPU/VRAM history storage, and fail-closed native fallbacks.

What it solves

This project provides a training-free acceleration method for the MiniMax H3 audio-video model within ComfyUI. It reduces the number of expensive transformer evaluations required during the sampling process, thereby speeding up video and audio generation without requiring additional training.

How it works

Based on the Spectrum method, the tool treats denoiser features as functions over diffusion time and approximates them using Chebyshev polynomial bases. It fits these coefficients online via ridge regression to forecast future feature states.

During an "actual step," the native MiniMax H3 model runs normally and captures the target hidden state. During a "forecast step," the tool predicts this hidden state from previous anchors and skips the transformer blocks entirely, while still running the native output and reconstruction path. For stochastic samplers like SEEDS, it forecasts only the transformer residual, adding it back to an exact current-state embedding to maintain stability.

Who it’s for

Users of ComfyUI who generate audio-video content using the MiniMax H3 model and want to reduce generation time while maintaining acceptable quality.

Highlights

  • Training-free acceleration: Speeds up sampling by skipping transformer evaluations without needing to be retrained.
  • Broad sampler support: Compatible with a wide range of samplers including Euler, ER-SDE, SEEDS-2/3, and SA-Solver PECE.
  • Integration with other tools: Works alongside ComfyUI-Untwisting-RoPE, Diff-Aid, and RefDelta Solver.
  • Approximate acceleration: Offers a flexible balance between speed and quality, with specific policies like balanced or max_speed for PECE forecasting.
  • H3-specific optimization: Specifically handles the packed audio-video representation of MiniMax H3.

Related

  • Dispatch
    MiniMax H3 Support in ComfyUI

    ComfyUI introduces day-zero support for MiniMax H3, an open-weights omni-modal video model capable of generating 2K video with native stereo audio on consumer hardware.

  • Project
    wildminder/ComfyUI-DyPE

    ComfyUI‑DyPE is a set of custom ComfyUI nodes that let you generate ultra‑high‑resolution images (2 K‑4 K+) with modern diffusion‑transformer models (Flux, Qwen‑Image, Z‑Image, etc.) without retraining. It provides training‑free patches (DyPE, SEGA, SPA, HAP) and cascade up‑scalers (PixelRush, FreeScale, HiFlow, I‑Max) that integrate as a single node after the model loader, offering quality, speed, or content‑preserving options.

  • Project
    Luisacaotica/ComfyUI-MiniMaxH3Mod

    ComfyUI‑MiniMaxH3Mod adds nodes that let you encode images, video, and audio into reusable `.safetensors` “RefMod” bundles for the MiniMax H3 model, load them, inspect token costs, and apply them as conditioning in ComfyUI pipelines.

  • Project
    Jasonzzt/ComfyUI-CacheDiT

    A ComfyUI extension that accelerates Diffusion Transformer (DiT) models through intelligent caching, providing 1.4x to 2.0x speedups for image and video generation.

  • Project
    HELPMEEADICE/TE-Speed-MiniMaxH3-OSS

    A ComfyUI custom node that accelerates MiniMax H3 model inference by up to 45% using a residual caching mechanism for DiT blocks.