ComfyUI Extension: ComfyUI-Node-for-Adaptive-Spectral-Feature-Forecasting-for-Diffusion-Sampling-Acceleration
Run ComfyUI workflows without the setup
No installs, no CUDA version roulette, no GPU sitting idle on your bill. Bring a workflow and run it in the browser.
A training-free ComfyUI model patcher that accelerates diffusion sampling via Chebyshev spectral forecasting, compatible with standard samplers like Euler and DPM++. (Description by CC)
README
Spectrum Patcher — ComfyUI Model Patch
Training-free diffusion acceleration via Chebyshev spectral forecasting (in the style of Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration).
Original Spectrum implementation/research repo: https://github.com/hanjq17/Spectrum
This custom node is a MODEL patcher: it wraps the model UNet/DiT forward via set_model_unet_function_wrapper(...) so it can work with normal ComfyUI samplers (Euler, DPM++, etc.).
Install
Clone into ComfyUI/custom_nodes.
Recommended (clone into a correctly-named folder):
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/maximilianwicen/ComfyUI-Node-for-Adaptive-Spectral-Feature-Forecasting-for-Diffusion-Sampling-Acceleration.git ComfyUI-Max-Spectrum
Then restart ComfyUI.
Node
- Menu name: Spectrum Patcher (Max)
- Category:
model/patches - Internal id:
MaxSpectrumPatcher
Parameters
window_size: run a real model pass every Nth step, forecast othersm: Chebyshev basis count / polynomial degreelam: ridge regularization strength for coefficient fittingw: blend factor between forecast and last real output on forecast steps
Notes
This implementation forecasts the model output tensor (noise/eps-style output) based on a short history of real evaluations. It is conservative by design: if the fit fails or produces NaNs/Infs, it falls back to a real model call.
Run ComfyUI workflows without the setup
No installs, no CUDA version roulette, no GPU sitting idle on your bill. Bring a workflow and run it in the browser.