Spectrum Adaptive Forecaster (SDXL)

A ComfyUI node in sampling with 9 inputs and 1 output.

By ruwwww·Created 5 months ago·Updated 3 months ago· 56

◄w0.30►

◄m3►

◄lam0.10►

◄window_size2►

◄flex_window0.00►

◄warmup_steps5►

◄stop_caching_step-1►

◄steps30►

Categorysampling

Inputs (9)

Name	Type	Default	Description
model	MODEL		The diffusion model to patch with Spectrum forecasting logic.
w	FLOAT	0.300–1	Blending weight between predicted (Chebyshev) and local (Taylor) features. Lower (0.3-0.5) preserves sharpness, higher relies on global smoothing.
m	INT	31–8	Number of Chebyshev basis functions (forecast complexity). Lower values (3-4) are more stable for SDXL.
lam	FLOAT	0.100–2	Ridge regularization strength. Prevents latent explosions and rainbow artifacts in low-precision modes.
window_size	INT	21–10	Initial forecasting window size (number of skipped steps).
flex_window	FLOAT	0.000–2	Increment added to window size after each actual UNet pass. Higher = aggressive acceleration.
warmup_steps	INT	50–20	Initial full-model steps before forecasting begins. Gives the model time to establish composition.
stop_caching_step	INT	-1-1–100	The exact step where Spectrum stops and returns to native UNet. Essential for final detail recovery. Set to Total Steps - 3.
steps	INT	3010–500	Match this value with your KSampler total steps for stable forecast accuracy and drift reduction.

Name	Type	Description
MODEL	MODEL	—