ComfyUI-Spectral-Preprocessing-Nodes

AdaptiveHFCompressor

Applies soft-knee dynamics compression (like an audio compressor) to the high-frequency band of the FFT magnitude. If the HF energy is below the threshold, nothing changes. Only excess energy above the threshold is reduced, following a configurable ratio and soft knee. The internal shape of the HF spectrum is preserved — only the overall level changes. | Parameter | Default | Description | |---|---|---| | threshold | 0.15 | RMS magnitude threshold. HF energy below this is untouched. | | ratio | 4.0 | Compression ratio. 4:1 = four units of excess become one unit. | | knee | 0.05 | Soft-knee width. 0 = hard knee. Larger = smoother transition. | | strength | 0.75 | Blend between original (0) and fully compressed (1). | | hfcutoff | 0.25 | Normalised radius above which signal is considered high-frequency. | | tilesize | 0 | Tile size for large images. | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 1 entrada · 7 parámetros · 1 salida

DirectionalArtifactSuppressor

Detects and attenuates frequency-domain energy that is anomalously concentrated along specific angular directions. Targets: checkerboards, ESRGAN stripe artifacts, CNN periodic patterns, aliasing bands, any periodic texture with a dominant orientation. How it works: Bins the FFT magnitude into angular sectors (0–180°). Estimates the smooth angular energy background via a circular median filter. Computes a per-angle z-score, builds a smooth sigmoid attenuation, maps it back to the 2-D Cartesian frequency grid. | Parameter | Default | Description | |---|---|---| | thresholdsigma | 3.0 | Z-score threshold for flagging a direction as anomalous. | | attenuation | 0.8 | Maximum fraction of energy removed in flagged directions. | | angularkernel | 5 | Half-width (bins) of the circular median filter. | | minradius | 0.05 | Minimum radius included in analysis. Pixels closer to DC are left alone. | | nanglebins | 180 | Angular resolution (bins over 0–180°). More bins = finer discrimination. | | tilesize | 0 | Tile size for large images. | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 1 entrada · 7 parámetros · 1 salida

FFTSpectrumVisualizer

Debug node: renders a 6-panel diagnostic grid of the FFT spectrum. Panels (left→right, top→bottom): 1. Log Magnitude — bright spots = concentrated spectral energy 2. Phase — uniform texture = natural; structured = artifacts 3. Radial Profile — white = measured, grey = 1/f² reference 4. Angular Profile — energy vs angle; spikes = directional artifacts 5. Spike Heatmap — z-score map from Node 1's detection step 6. Gain Mask — what Node 1 would apply at these parameters Connect the output to a PreviewImage node to inspect it.

Spectral Preprocessing · 1 entrada · 5 parámetros · 1 salida

HermitianSymmetryEnforcer

Restores the Hermitian symmetry of the FFT of a real-valued image. For any real image I, the DFT must satisfy F[k,l] = conj(F[−k,−l]). This constraint can be subtly violated by lossy codec cycles, certain interpolation methods, or manual frequency-domain edits. Violations cause the IFFT to produce a non-trivial imaginary component that is silently discarded by every .real call — losing spectral energy that should be in the image. This node corrects the spectrum to the nearest Hermitian matrix (minimum L2): F_corrected = ( F + conj(F_flipped) ) / 2 The correction is exact and deterministic. For clean natural images the residual is negligible (< 1e-6 in pixel values). For images that have been through multiple codec rounds or manual FFT edits, it can be visible. The second output "residual" shows the imaginary part of IFFT(original spectrum) — use it as a diagnostic; nearly-black = no problem. Recommended placement: after any node that performs manual magnitude / phase editing, or as a "clean-up" step before VAE encoding.

Spectral Preprocessing · 1 entrada · 1 parámetro · 2 salidas

LogSpectrumPeakCompressor

Processes the FFT magnitude in the logarithmic domain, inspired by cepstral and homomorphic signal processing. Working in log space turns multiplicative spectral distortions (gain patterns, grid artifacts, repetitive texture overlays) into additive signals that are much easier to separate and compress. Algorithm: FFT → log(magnitude) → smooth background → compress residual peaks → exp → IFFT. The phase is never touched. | Parameter | Default | Description | |---|---|---| | peakthreshold | 1.0 | Log-domain residual threshold. Smaller = more aggressive. | | ratio | 4.0 | Compression ratio applied to log-domain peaks above the threshold. | | smoothing | 0.05 | Background blur strength as fraction of image size. | | strength | 0.8 | Blend between original (0) and fully compressed (1). | | backgroundmode | gaussian | gaussian = smoother envelope; uniform = faster. | | tilesize | 0 | Tile size for large images. | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 1 entrada · 7 parámetros · 1 salida

MoisanDecomposition

Removes the FFT cross artifact via Moisan's Periodic+Smooth Decomposition. The "cross" (bright horizontal + vertical lines through DC in any FFT spectrum) is caused by the non-periodicity of the image at its borders. Moisan's algorithm separates an image into: • u (periodic component) — spectrally clean, no cross artifact. • v (smooth component) — the boundary discontinuity field. The main output is u (strength=1) or a blend toward the original image (strength < 1). The smooth component v is available as a second output for optional recombination after downstream spectral processing: result = SpectralNodes(u) + v This node has no trainable parameters — the decomposition is exact and deterministic. The only free parameter is strength, which controls how aggressively the smooth component is subtracted. Recommended placement: first node in any FFT preprocessing chain.

Spectral Preprocessing · 1 entrada · 1 parámetro · 2 salidas

NoiseFloorLifter

Estimates the uniform spectral noise floor (the background energy level in the FFT magnitude) and subtracts it. Analogous to spectral subtraction in audio processing. Particularly effective on AI-generated images and heavily compressed sources where the noise floor is elevated relative to natural photographs. Because the subtraction is global and uniform, it introduces zero spatial ringing. | Parameter | Default | Description | |---|---|---| | floorpercentile | 10.0 | Percentile of magnitude distribution used as the noise floor estimate. | | strength | 0.7 | Fraction of the noise floor to subtract. | | oversubtraction | 1.0 | Multiplier on floor estimate. Values > 1 are more aggressive. | | preservedc | True | Leave the DC coefficient (mean brightness) untouched. | | tilesize | 0 | Tile size for large images. | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 1 entrada · 6 parámetros · 1 salida

RadialSpectrumNormalizer

Natural photographs follow a power law: Power(f) ∝ 1/f^α with α ≈ 2. AI-generated images often carry excess high-frequency energy. This node estimates the current radial power spectrum, builds a smooth correction gain curve toward the target law, and blends it in. Left: source with injected high-frequency noise (excess HF visible as a bright haze at the edges of the FFT spectrum). Right: after Radial Spectrum Normalizer — HF energy reduced, spectrum closer to the natural 1/f² slope. | Parameter | Default | Description | |---|---|---| | alpha | 2.0 | Exponent of target 1/f^alpha power law. Natural photos ≈ 2.0. | | strength | 0.5 | Blend factor. 0 = no change, 1 = full correction. Start low. | | preservelowfreq | 0.1 | Normalised radius below which gain is frozen at 1.0. | | smoothing | 0.05 | Gaussian smoothing of the radial gain curve. Higher = fewer ringing artifacts. | | tilesize | 0 | Tile size for large images. | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 1 entrada · 6 parámetros · 1 salida

RadialStratifiedHistogramMatch

The frequency-domain equivalent of CLAHE — performs histogram matching of FFT magnitudes independently within each radial frequency band. A global CDF conflates low- and high-frequency coefficients that follow completely different statistical distributions. Per-band matching respects this: each annular ring gets its own CDF mapping, giving accurate per-scale texture transfer. This is the most expressive spectral matching node in the pack — it simultaneously matches the spectral statistics at every frequency scale independently, requiring no assumptions about the target curve shape (unlike RadialSpectrumNormalizer's 1/f^α model). | Parameter | Default | Description | |---|---|---| | strength | 0.75 | 0 = no change, 1 = fully match target per-band distribution. | | nbands | 16 | Number of radial annuli. 8–24 is a good range. | | preservelowfreq | 0.05 | Radius below which bands are skipped — protects global brightness and colour. Raise to 0.15–0.2 for pre-processing. | | tilesize | 0 | Tile size (leave at 0 for global statistics — tiling changes band statistics). | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 2 entradas · 5 parámetros · 1 salida

SpectralAngularEqualizer

Equalizes systematic directional biases in the FFT spectrum. Unlike DirectionalArtifactSuppressor (which targets isolated spikes), this node normalizes the *entire* angular energy distribution toward the isotropic mean. It corrects for: • Horizontal/vertical texture bias (architecture, screens, fabric) • Scanner line artifacts • JPEG block-grid artifacts (energy peaks at 0° and 90°) • AI-generation directional frequency preferences The equalize_power parameter controls aggressiveness: • 1.0 = full normalization — every angle gets the same mean energy • 0.5 = half-way correction — softer, preserves more of the original bias • 0.0 = no effect (identity) Gains are clamped and the gain curve is smoothed across angles to avoid sharp discontinuities between adjacent bins.

Spectral Preprocessing · 1 entrada · 8 parámetros · 1 salida

SpectralChannelEqualizer

Reduces inter-channel spectral mismatch to suppress colour fringing artifacts. Computes the radial power spectrum per channel, takes the mean across R/G/B as a reference, and applies a smooth radial gain per channel to equalise them. Global colour balance is protected by preservelowfreq. Targets: colour fringing, chromatic halos, colour noise — common in AI-upscaled and AI-generated images. | Parameter | Default | Description | |---|---|---| | strength | 0.5 | How strongly to equalise channel spectra. 0 = no change, 1 = full equalization. | | preservelowfreq | 0.08 | Normalised radius below which channel gains are frozen at 1.0. | | smoothing | 0.06 | Smoothing sigma for radial gain curves. | | tilesize | 0 | Tile size for large images. | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 1 entrada · 5 parámetros · 1 salida

SpectralHistogramMatch

Matches the statistical distribution of FFT coefficient magnitudes from imagesource to imagetarget — the frequency-domain analogue of histogram matching. Unlike pixel-space histogram matching, colour and spatial structure are not affected — only the distribution of spectral energy changes. Phase is always preserved, so edges, object positions, and generated content survive intact. Pre-processing use: make a synthetic or AI-generated image spectrally indistinguishable from a natural photograph before VAE encoding. Post-processing use: give a generated image the same grain and texture character as the source, without altering generated colours or content. Left to right: natural photo (target), AI-like source (excess HF noise + brightness shift), after Global Histogram Match, after Radial Stratified Histogram Match. FFT spectra shown below each image — note how the matched spectra converge toward the target. | Parameter | Default | Description | |---|---|---| | strength | 0.75 | 0 = no change, 1 = fully match target magnitude distribution. | | tilesize | 0 | Tile size for large images. 0 = whole image. | | tileoverlap | 64 | Tile overlap in pixels. |

Spectral Preprocessing · 2 entradas · 3 parámetros · 1 salida

SpectralMagnitudeBlend

Blends the FFT magnitudes of two images while preserving the phase (spatial structure) of image A. The magnitude encodes what kind of texture; the phase encodes where things are. Keeping phaseA and blending magnitudes produces an image that retains the spatial layout of A but gradually adopts the spectral character of B. Typical use: feed a known-clean natural photograph as imageb and an AI-generated image as imagea. At blend=0.3–0.5 the output has A's content but B's spectral texture, making the AI image feel more natural to the VAE. | Parameter | Default | Description | |---|---|---| | blend | 0.3 | 0 = keep A's magnitude entirely, 1 = use B's magnitude entirely. | | frequencyband | all | Restrict blending to all, low, mid, or high frequency band. |

Spectral Preprocessing · 2 entradas · 2 parámetros · 1 salida

SpectralPhaseBlend

Blends the FFT phases of two images while retaining the magnitude of image A. Phase encodes spatial structure — blending phases morphs the spatial geometry of the image while its spectral character is preserved. Uses circular interpolation (angle(lerp(exp(iφa), exp(iφb), t))) to correctly handle the ±π phase wrap. ⚠ Even small blend values (0.05–0.1) produce visible structural changes. Start low. | Parameter | Default | Description | |---|---|---| | blend | 0.05 | 0 = keep A's phase, 1 = use B's phase. Start at 0.05. | | frequencyband | all | Restrict blending to all, low, mid, or high frequency band. |

Spectral Preprocessing · 2 entradas · 2 parámetros · 1 salida

SpectralSpikeSuppressor

Detects and attenuates isolated peaks in the FFT magnitude spectrum. Targets periodic artifacts that produce narrow spikes in frequency space: JPEG ringing, checkerboard patterns from transposed convolutions, Moiré fringes, CNN grid artifacts. How it works: FFT → median-filter background → MAD-based z-score map → smooth sigmoid attenuation mask → reconstruct using original phase → IFFT. Phase is never modified, so edges are preserved exactly. | Parameter | Default | Description | |---|---|---| | thresholdsigma | 3.0 | Z-score above which a coefficient is treated as a spike. Lower = more aggressive. | | attenuation | 0.85 | Maximum fraction of spike energy removed. 1.0 = full suppression. | | kernelsize | 15 | Median filter size for background estimation. Larger = more context. | | preservedc | True | Protect the DC coefficient (mean brightness). Recommended: on. | | tilesize | 0 | Tile size for large images. 0 = whole image at once. | | tileoverlap | 64 | Overlap between tiles in pixels. |

Spectral Preprocessing · 1 entrada · 6 parámetros · 1 salida

SpectralWhitening

Normalizes the FFT spectrum by the measured radial power envelope. Unlike RadialSpectrumNormalizer (which fits a 1/f^α model), this node is fully data-driven — it divides each coefficient by the ACTUAL measured power at its radial frequency, with no model assumption. This corrects any radial bias, regardless of its shape. Two modes: • "flatten" — true whitening: all radial bands get equal mean power. • "target_1f" — data-driven reshape to target_alpha slope (default 2.0 = natural pink noise). More aggressive than RadialSpectrumNormalizer because it measures and corrects the actual bias rather than assuming a model shape. Phase is never modified. The DC coefficient is always protected.

Spectral Preprocessing · 1 entrada · 9 parámetros · 1 salida