Radial Spectrum Normalizer

Natural photographs follow a power law: on average, spectral power falls off as 1/f^α with α close to 2. AI-generated images - especially over-sharpened ones and anything that's been through an eager upscaler - carry excess high-frequency energy, so their spectrum sits above the natural curve in the highs. You can't see that excess in the image; it just makes the output read as "synthetic" and makes the VAE encoder fidget. RadialSpectrumNormalizer is the node in ComfyUI-Spectral-Preprocessing-Nodes that measures your spectrum's actual falloff and gently bends it back toward the natural one.

It's the model-based sibling of SpectralWhitening (which measures and corrects the actual radial bias without assuming a curve shape). This node assumes the 1/f^α law and fits toward it, which makes it more robust when the image is mostly healthy, and less aggressive than the data-driven alternative.

How it works

The node estimates the current radial power spectrum - mean energy at each distance from the DC center - then builds a smooth correction gain curve that would reshape it toward the target 1/f^alpha law, and blends that correction in with strength. Three safety rails keep it sane: preserve_low_freq freezes gain at 1.0 inside a small radius (so global structure and brightness are untouched), smoothing Gaussian-smooths the gain curve (so you don't get per-ring ringing), and the gain is blended rather than fully applied. Phase is never touched.

The inputs that matter

• alpha (default 2.0) - the exponent of the target law. Natural photos ≈ 2.0; higher pushes more energy toward low frequencies (darker, smoother); lower flattens the slope.
• strength (default 0.5) - how much of the correction to apply. The README explicitly says start low - 0.3–0.6 is its recommended band for "too synthetic" texture.
• preserve_low_freq (default 0.1) - normalised radius below which gain is frozen at 1.0. Protects global structure; leave near default.
• smoothing (default 0.05) - Gaussian smoothing sigma on the gain curve. Higher = smoother transition, less ringing, at the cost of a looser correction.

Then the pack-standard tile_size / tile_overlap for large images. One image output.

Installing it

Part of ComfyUI-Spectral-Preprocessing-Nodes - one install, sixteen nodes, all under Spectral Preprocessing. ComfyUI Manager (search "ComfyUI-Spectral-Preprocessing-Nodes"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/EdoardoGuerriero/ComfyUI-Spectral-Preprocessing-Nodes

Restart. Dependencies are numpy and scipy, both shipped with ComfyUI. Nothing else to fetch.

Where people get burned

The trap is treating it as a sharpness dial. It isn't - it's a statistical correction, and if you push strength to 1.0 the image can lose texture character even though no pixel-level blur was applied. The other classic miss: it's the wrong tool when the problem isn't a smooth HF excess but discrete artifacts (checkerboards, spikes) - that's the Spike Suppressor's job. The README's own demo shows the target: take the source with injected HF noise, and the after-image's spectrum settles onto the grey 1/f² reference line. Run the FFT Spectrum Visualizer's radial-profile panel first: if your spectrum sits above the grey reference in the highs, this node is the fix; if it's roughly on the line already, you're over-processing and should leave it out. When in doubt, strength 0.3 and a glance at the profile beats guessing.