ComfyUI-Spectrum-MiniMax-H3

Why we included this project

Sampling MiniMax H3 in ComfyUI can be slow, and this custom node speeds it up by skipping some of the model's transformer evaluations. It's a training-free accelerator that fits a Chebyshev ridge regression to the hidden state after the final transformer block, then uses those forecasts on later solver steps while still running the native output heads and video/audio reconstruction each step. The speedup is approximate, not lossless, so output can differ from native H3 even with the same seed and workflow; the project is honest about that trade-off. Safeguards include adaptive scheduling, sampler-aware fail-closed behavior, CPU or VRAM history storage, and native fallbacks when a forecast can't be trusted. For people running text-to-video, first/last-frame, or reference-conditioned MiniMax H3 workflows on constrained GPUs, the bounded memory controls and explicit compatibility checks that fail loudly if ComfyUI changes break things are a real plus, and you'll want to A/B test quality against the native path.