Frequency-Decoupled Guidance

Most of the ways you'd normally sharpen up an image - raising CFG, adding PAG on top - buy you detail at the cost of oversaturation and burn once you push far enough. Frequency-Decoupled Guidance takes a different angle: it's an implementation of "Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales," which is exactly what it sounds like - split the guidance signal by frequency and let low CFG do the heavy lifting instead of fighting it.

The core idea, per the paper's title, is that image structure (the low-frequency stuff - layout, big shapes, color blocks) and fine detail (high-frequency - texture, edges, fine noise) don't need the same guidance treatment. Standard CFG applies one guidance strength to everything, which is part of why pushing it hard for more detail also over-guides the coarse structure and burns the image. Decoupling the two lets you keep detail crisp while running your actual CFG scale low.

This is the barest node in the whole pack, surface-wise: one knob.

The input that matters

• strength_high (default 12, range 0–100) - guidance strength for the high-frequency band, i.e. fine detail and texture. There's no exposed low-frequency counterpart; the coarse-structure guidance rides on your sampler's regular CFG as usual. So the way you'd actually use this: keep CFG modest (per the paper's whole premise), and use strength_high to dial in how much extra guidance goes toward fine detail specifically.

Output is a single MODEL, same as every other node in this pack - wire it into your KSampler.

How to install it

ComfyUI Manager: search "Perturbed-Attention Guidance" (this node ships inside that pack, not as a separate registry entry). Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/pamparamm/sd-perturbed-attention

Restart ComfyUI afterward. No model downloads, no extra Python dependencies to fight with.

Common issues & troubleshooting

Straight talk: this is one of the newest, most obscure nodes in the pack. It's a fresh implementation of a 2025 paper with essentially no forum trail yet - nobody's written the "here's what I learned tuning this" post, so there's no accumulated community wisdom to pass along, and I'm not going to invent failure modes that haven't actually been reported.

What you can do sensibly: treat the default of 12 as a starting point, not gospel. Fix your seed, change only strength_high, and compare - the same discipline that applies to tuning CFG itself, since single-image before/afters on a random seed prove nothing about whether a change actually helped or you just got a luckier roll.

If you're stacking this with other guidance patches from this pack (PAG, NAG, SEG), remember they're all just MODEL → MODEL patches chained in sequence - if the combination goes sideways, pull them out one at a time rather than guessing which interaction is the culprit. Given how sparse the field reports are on this specific node, isolating it first is the fastest way to actually learn something about what it's doing on your workflow.