ComfyUI-Spectrum-sdxl

ComfyUI Spectrum SDXL Node

This repository contains a Unnoficial ComfyUI custom node implementing the Spectrum sampling acceleration technique, tailored specifically for SDXL models. Spectrum is a training-free method that forecasts spectral features using Chebyshev polynomials and ridge regression to skip redundant UNet computations, achieving significant speed-ups with less quality loss.

Disclaimer: This implementation follows the official Spectrum paper logic but caches the outer feature rather than the inner feature approach.

Legacy / Non-Faithful Node: The SpectrumSDXLCalibrated node is now considered legacy.

Clarification for returning users: The previous version of this node was essentially "vibe-coded" from scratch because I couldn't initially find the official forecaster implementation. This led to some non-principled additions like "calibration" which, while interesting, are not faithful to the paper. This has now been solved by porting the official forecaster code into the SpectrumSDXL node. Please migrate to the faithful implementation for more stable and principled results.

Performance comparison — Download the images below and drag them into ComfyUI to instantly load optimized workflows!

SDXL Calibrated Comparisons (30-step Euler)

Normal	Spectrum	Calibrated (strength 0.5)	Calibrated (strength 0.8)
Normal	Spectrum	Cal0.5	Cal0.8
8.8 s	4.8 s	4.8 s	4.8 s
Normal2	Spectrum2	Cal0.5_2	Cal0.8_2
8.3 s	4.6 s	4.6 s	4.6 s

Anima (30-step er-sde)

Default	Spectrum
Default Anima	Spectrum Anima
23.67 s	13.01 s

---

Key Features

• Sampling Acceleration – Reduce inference time (up to ~2× on SDXL) by skipping UNet evaluations on selected timesteps.
• Vectorized Batch Processing – Fully vectorized mathematical operations process conditional and unconditional latents independently. This prevents memory contamination (rainbow artifacts) and sustains ultra-high it/s speeds without Python loop bottlenecks.
• FP8 Tensor-Core Support – Compatible to run on NVIDIA Tensor Cores in FP8 mode, providing additional speed gains on compatible hardware. Works seamlessly alongside other optimizations.
• Sage-Attention Friendly – Orthogonal to Sage Attention; you can enable them together without conflicts.
• Stability Enhancements – Incorporates a "Sliding Window" memory limit (prevents mathematical explosions), alongside jitter and anti-NaN safeguards to keep low-precision runs (FP16/FP8) perfectly stable.
• Step-Based Quality Guard – A dedicated parameter allows you to seamlessly turn off the forecaster during the final critical steps, preserving high-frequency micro-details (like skin texture and eyes) and eliminating VAE-like compression blur.

Installation

1. Navigate to the custom_nodes directory of your ComfyUI installation.
2. Clone this repository:

git clone https://github.com/ruwwww/comfyui-spectrum-sdxl

3. Restart ComfyUI.

Parameters

| Parameter | Description