ComfyUI Custom Node — Phase Congruency Edge

A ComfyUI custom node that detects edges and features using Phase Congruency, an illumination-invariant and contrast-invariant feature detector based on the coherence of Fourier phase components across multiple scales and orientations (Kovesi, 1999).

---

What is Phase Congruency?

Most classical edge detectors (Sobel, Canny, Laplacian) rely on intensity gradients, so their output changes with lighting conditions and local contrast. Phase Congruency takes a fundamentally different approach:

Features are detected at locations where Fourier components are maximally in phase with one another — regardless of their amplitude.

This means the detector is:

• Illumination invariant — absolute brightness has no effect
• Contrast invariant — works equally on high-contrast and low-contrast regions
• Perceptually meaningful — closely matches how the human visual system perceives edges and corners
• Theoretically grounded — based on the Kovesi (1999) frequency-domain formulation using log-Gabor filters

---

Node

Property	Value
Node name	Phase Congruency Edge
Category	image/filters
Input	IMAGE (RGB or grayscale, any resolution)
Output	IMAGE (grayscale edge map, white = strong edges)

---

Parameters

Parameter	Type	Default	Range	Description
image	IMAGE	—	—	Input image tensor
nscale	INT	4	2–8	Number of log-Gabor filter scales
norient	INT	6	2–8	Number of filter orientations
min_wavelength	INT	3	2–20	Minimum wavelength of the finest-scale filter (pixels)
mult	FLOAT	2.1	1.5–4.0	Multiplicative factor between successive filter scales
sigma_on_f	FLOAT	0.55	0.1–1.0	Bandwidth of each log-Gabor filter (ratio σ/f₀)
k	FLOAT	2.0	0.5–10.0	Noise threshold in standard deviations above the mean
cutoff	FLOAT	0.5	0.1–0.9	Butterworth low-pass filter cutoff frequency

---

Algorithm Overview

For each of the norient orientations:

1. Build a log-Gabor bandpass filter at each of the nscale scales, multiplied by an angular spread function (raised cosine) centered on the current orientation.
2. Apply each filter in the frequency domain (via FFT) to obtain complex-valued responses.
3. Sum the complex responses across all scales to form an energy vector (E_real, E_imag).
4. Estimate the noise threshold T using Rayleigh statistics on the finest-scale amplitude:

   T = mean_noise + k × noise_std
   

5. Compute per-orientation phase congruency:

   PC_orient = max(|energy| − T, 0) / (Σ amplitude + ε)
   

Final phase congruency map = mean of PC_orient across all orientations, normalized to [0, 1].

---

Parameter Tuning Guide

Goal	Adjustment
Detect finer / more detailed edges	Decrease min_wavelength
Detect coarser / broader edges	Increase min_wavelength
More sensitive (at risk of noise)	Decrease k (e.g. 1.0)
Cleaner output, fewer false edges	Increase k (e.g. 4.0–6.0)
Improve isotropy (all directions)	Increase norient (e.g. 8)
Capture more scale information	Increase nscale (e.g. 6)
Reduce ringing / high-freq artefacts	Decrease cutoff (e.g. 0.4)

Recommended starting point: nscale=4, norient=6, min_wavelength=3, mult=2.1, sigma_on_f=0.55, k=2.0

---

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/bemoregt/ComfyUI_CustomNode_PhaseCongruencyEdge.git

Then restart ComfyUI. The node will appear under image/filters → Phase Congruency Edge.

Dependencies

Only standard packages are required — no extra pip install step needed if ComfyUI is already running:

numpy
pillow
torch

---

Comparison with Other Edge Detectors

Method	Illumination Invariant	Contrast Invariant	Multi-scale	Notes
Sobel	No	No	No	Fast, simple gradient
Canny	No	Partial	No	Needs manual threshold
Laplacian of Gaussian	No	No	No	Sensitive to noise
Phase Congruency	Yes	Yes	Yes	Perceptually meaningful

---

Example Use Cases

• Preprocessing for segmentation pipelines
• Artistic / stylized edge rendering
• Document scanning and OCR preprocessing
• Medical image feature extraction
• Texture analysis

---

Reference

Kovesi, P. (1999). Image Features from Phase Congruency. Videre: Journal of Computer Vision Research, 1(3).

---

License

MIT License — free for personal and commercial use.