Spectral Whitening

Spectral whitening is the classic signal-processing move: measure the actual power at every frequency, then divide each coefficient by its measured power so every band ends up equally loud - the spectrum is "whitened." SpectralWhitening brings that to ComfyUI, but with a twist that makes it genuinely different from the pack's other radial node: it's fully data-driven, with no model assumption about what the spectrum "should" look like.

Compare it to RadialSpectrumNormalizer, which fits the 1/f^α law and nudges your spectrum toward it. That node assumes the curve. This one measures your image's actual radial power envelope and corrects whatever bias is really there - a bump, a hump, an unexpected shape - regardless of its form. The README positions it as more aggressive than RadialSpectrumNormalizer precisely because it corrects the actual measured bias rather than assuming a model shape. If your image has a weird spectral signature that no power-law model would capture, this is the node that can handle it.

How it works

The node bins the FFT magnitude into n_radial_bins rings, measures the mean power in each ring, and builds a gain curve that's the inverse of that envelope - dividing each coefficient by the actual power at its radius. Two modes:

• flatten - true whitening: every radial band gets equal mean power. The source calls this aggressive, because a fully flat spectrum is not what natural images look like.
• target_1f (default) - data-driven reshape toward a target_alpha slope (default 2.0, natural pink noise). Same measurement-and-correct mechanism, but aimed at the natural curve instead of full flatness. This is the sane default.

Safety rails: the DC coefficient is always protected, gains are clamped by min_gain/max_gain (so the noise floor can't be amplified without limit and near-DC bands can't be crushed), and smoothing Gaussian-smooths the gain curve to prevent per-ring ringing. Phase is never touched.

The inputs that matter

• mode (default target_1f) - target_1f is the practical choice; flatten is the science experiment.
• target_alpha (default 2.0) - target slope for target_1f mode; 2.0 = natural pink noise. Ignored in flatten mode.
• strength (default 0.6) - blend between original and fully whitened. This is your main dial; the pack's style is "start low and verify."
• n_radial_bins (default 64) - radial resolution for power estimation. The tooltip's sweet spot is around 64.
• smoothing (default 0.08) - smoothing sigma for the gain curve, as a fraction of n_radial_bins; 0.05–0.15 recommended.
• min_gain / max_gain (defaults 0.05 / 20) - the clamps. Leave them.

Plus the pack-standard tile_size/tile_overlap. One image output.

Installing it

Part of ComfyUI-Spectral-Preprocessing-Nodes - one install, all sixteen nodes 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. Only numpy and scipy (bundled); nothing to download.

Where people get burned

Two things. First, flatten mode is a trap for anyone who reads "whitening" and thinks "yes, maximum." A flat spectrum is not a natural image, and full whitening on strength 1.0 will make the output look odd and textureless - keep mode on target_1f unless you specifically want to see what flat looks like. Second, more aggressive than its model-based sibling means it can over-correct if the per-bin statistics are noisy: that's what smoothing is for, so don't drop it below ~0.05 on small images. The honest guidance from the pack's own framing: use RadialSpectrumNormalizer when your image is mostly healthy and you just want a gentle nudge toward natural; reach for SpectralWhitening when the spectrum has a shape no 1/f^α model explains. When in doubt, target_1f, strength 0.3–0.6, and the visualizer's radial-profile panel will tell you the rest.