Spectral Angular Equalizer

Some images have a directional bias you can feel before you can see: scanner line artifacts, JPEG block-grid energy piling up at 0° and 90°, fabric or architecture texture leaning horizontal-vertical, AI generators quietly preferring certain orientations. DirectionalArtifactSuppressor in this pack hunts isolated spikes in specific directions; SpectralAngularEqualizer does something broader and more systematic - it normalizes the entire angular energy distribution toward the isotropic mean, so no direction is dramatically louder than the others.

It's the angular counterpart to SpectralWhitening's radial equalization: where the radial nodes balance energy across frequency scales, this one balances energy across orientations. The source lists its targets bluntly: horizontal/vertical texture bias, scanner lines, JPEG block-grid artifacts, and AI-generation directional frequency preferences.

How it works

The FFT magnitude is binned into n_angle_bins angular sectors across 0–180°. For each sector, the node computes a gain that would pull that sector's mean energy toward the global mean. The equalize_power exponent controls how hard: at 1.0 every angle gets equalized to the same mean energy; at 0.5 you get a square-root gain - half the correction, softer, preserving more of the original bias; at 0.0 it's identity. Gains are clamped by min_gain/max_gain (so a near-silent direction can't be amplified into a 100× noise booster, and a dominant direction can't be crushed to nothing) and smoothed across adjacent bins so there are no sharp angular discontinuities. Phase untouched, and coefficients inside protect_radius are excluded entirely.

The inputs that matter

• equalize_power (default 0.75) - the aggressiveness exponent, and the first knob to turn. The tooltip recommends starting at 0.5–0.75.
• n_angle_bins (default 36) - angular resolution (36 = 5° per bin). The tooltip's suggested range is 36–72; more bins = finer but noisier statistics.
• protect_radius (default 0.08) - normalized radius below which coefficients are excluded. This is the important safety valve: the source notes that low-frequency anisotropy is often real image content (a landscape is legitimately not isotropic at low frequencies), so it recommends 0.05–0.15.
• min_gain (default 0.1) - the floor on per-angle gain. At 0.1, the quietest direction can be boosted by at most 10×.
• max_gain (default 10.0) - the ceiling, preventing extreme suppression of dominant directions.
• strength (default 0.7) - overall blend between original and fully equalized.

Plus the pack-standard tile_size/tile_overlap for large images. 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 with ComfyUI); nothing to download.

Where people get burned

The failure mode is over-correcting real directionality into mush. An architectural shot should have strong vertical and horizontal energy - that's content, not artifact - and only protect_radius shields the low frequencies from your equalization. If you push equalize_power toward 1.0 on such an image, you'll start flattening that real structure. The useful differentiator the pack itself draws: use DirectionalArtifactSuppressor when you can point at a specific artifact direction, use this node when the bias is systemic across many angles (scanner grain, block grids). And check the visualizer's angular-profile panel before committing - a wide, smooth asymmetry is this node's target; a single needle is the suppressor's.