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 native ComfyUI implementation of Spectrum, a training-free method designed to accelerate the sampling process of the MiniMax H3 audio-video model. It reduces the number of expensive transformer evaluations required to generate media, thereby speeding up the generation process without requiring additional training.

How it works

Spectrum treats denoiser features as functions over diffusion time and approximates them using Chebyshev polynomial bases. It uses ridge regression to fit these coefficients online, allowing it to forecast future feature states.

In a typical generation workflow, Spectrum alternates between "actual" steps (where the native MiniMax H3 transformer runs normally) and "forecast" steps. During a forecast step, the model predicts the target hidden state from previous actual anchors and skips the expensive transformer blocks, while still running the native output and sampler path. This is an approximate accelerator, meaning the output may differ slightly from a native H3 run.

Who it’s for

Users of ComfyUI who are generating audio-video content using the MiniMax H3 model and want to reduce sampling time and compute costs.

Highlights

  • Training-Free Acceleration: Accelerates MiniMax H3 without needing to model weights updates.
  • Broad Sampler Support: Compatible with a wide range of samplers, including Euler, ER-SDE, SEEDS-2/3, and SA-Solver PEC/PECE.
  • Integration with H3 Ecosystem: Works with other H3-specific tools like RefDelta Solver, ComfyUI-Untwisting-RoPE, and H3 Continuum.
  • Flexible Forecasting Policies: Offers different forecast policies (e.g., balanced or max_speed) to balance speed and quality.
  • State-Conditioned Residuals: For stochastic samplers like SEEDS, it forecasts only the transformer residual to maintain noise process integrity.

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
    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.

  • Project
    LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

    A custom ComfyUI node set for Minimax H3 that provides neural latent upscaling and tiled re-sampling to accelerate high-resolution video generation while avoiding VAE round-trip overhead.

  • Project
    NikoDemon80/ComfyUI-H3-Motion-Context

    A ComfyUI extension for MiniMax H3 that enables seamless chaining of video and audio clips by using latent-based continuity to prevent quality loss and motion and sound jumps.

  • 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.