ComfyUI-Spectrum

A custom ComfyUI node that brings a Spectrum-style acceleration path for FLUX into a real ComfyUI workflow.

This node is designed to sit directly on the MODEL path in a FLUX workflow:

UNETLoader -> LoRA(s) -> Spectrum Apply Flux -> CFGGuider

What it does

This custom node patches the internal FLUX model used by ComfyUI and applies a pre-head feature forecasting strategy inspired by Spectrum.

Instead of forecasting the final denoised output directly, it forecasts an internal image feature right before the final FLUX head (final_layer), then still applies the original head with the current conditioning. In practice, this gave much better image stability than forecasting the model output directly.

Current scope

• Backend: FLUX only
• Integration target: ComfyUI FLUX workflows
• Tested setup: ComfyUI + FLUX + LoRA stack + CFGGuider + RES4LYF
• Node type: MODEL -> MODEL

This is not a universal Spectrum port for every diffusion backend. It is a targeted ComfyUI implementation for FLUX.

Why this exists

The official Spectrum repository reports large speedups in its own pipelines. In ComfyUI, the internal FLUX execution path is different enough that a direct drop-in port is not practical.

This node adapts the idea to ComfyUI by:

• patching Flux.forward_orig
• forecasting the pre-head image feature
• keeping the original final_layer
• synchronizing decisions with the sigma schedule from transformer_options

That makes it much more stable than a naive block-skipping or output-forecasting approach.

Installation

1. Download this repository or unzip the release.

2. Copy the folder into:

   ComfyUI/custom_nodes/ComfyUI-Spectrum/

3. Restart ComfyUI.

Node included

Spectrum Apply Flux

This is the only production node in the package.

It takes a MODEL input and returns a patched MODEL.

Place it:

• after your FLUX model loaders / LoRA stack
• before CFGGuider

Example:

UNETLoader -> LoRA Loader(s) -> Spectrum Apply Flux -> CFGGuider

Recommended settings

These defaults were the best working settings in the tested workflow:

• m = 4
• lam = 0.10
• mix_w = 0.50
• window_size = 2.0
• flex_window = 0.75
• warmup_steps = 0
• k_history = 16
• debug = false

Observed result in the tested workflow

In the tested setup, the best result observed was approximately:

• baseline: ~160 s
• with Spectrum Apply Flux: ~90 s

That is about a 44% speedup while keeping the image visually coherent in the tested cases.

Your results may vary depending on:

• FLUX variant
• sampler
• scheduler
• step count
• LoRA stack
• image-to-image vs text-to-image
• other custom nodes that wrap the model

Limitations

• FLUX only
• tuned for a specific ComfyUI execution path
• not guaranteed to match the exact speedup figures from the official Spectrum repository
• not guaranteed to behave the same on every sampler or custom extension
• this is still an adaptation, not an official upstream implementation

Troubleshooting

The node appears but does nothing

Make sure it is inserted on the MODEL line, not on latent, conditioning, or image paths.

Correct placement:

UNETLoader -> LoRA(s) -> Spectrum Apply Flux -> CFGGuider

The image quality collapses

Use the recommended defaults first. More aggressive settings are not always faster or better in real ComfyUI workflows.

It is slower than expected

That can happen if:

• your step count is very low
• your sampler interacts with the model in a way that reduces the benefit
• startup tasks or background fetches are happening during the benchmark
• your workflow differs significantly from the tested one

Credits

• Inspired by the official Spectrum project: https://github.com/hanjq17/Spectrum
• Adapted for ComfyUI FLUX workflows

First custom node

If this is your first custom node: nice one. This project went through several iterations before reaching a stable and useful result. The current version exists because the earlier, simpler approaches were tested against real ComfyUI behavior and then replaced with a more faithful pre-head strategy.

License

No license file is included by default in this package. Add one before publishing publicly if you want others to reuse or modify the code under clear terms.