ComfyUI Phase Stretch Transform Node

A ComfyUI custom node that applies the Phase Stretch Transform (PST) to images. PST is a physics-inspired edge detection algorithm that enhances features by simulating wave propagation through a dispersive medium.

Installation

1. Clone or copy this folder into your ComfyUI custom_nodes directory:

   ComfyUI/custom_nodes/Comfyui_PST/
   

2. Install dependencies:

   pip install numpy scipy
   

3. Restart ComfyUI.

The node will appear under image → processing in the node menu.

Node: Phase Stretch Transform

Inputs

| Name | Type | Default | Description | |---|---|---|---| | image | IMAGE | — | Input image (RGB or grayscale) | | lpf_sigma | FLOAT | 0.21 | Gaussian low-pass filter sigma applied before PST. Higher values smooth the image more, reducing noise sensitivity. | | phase_strength | FLOAT | 0.48 | PST phase strength (S). Controls how strongly phase variations are amplified. | | warp_strength | FLOAT | 12.14 | PST warp strength (W). Controls the frequency-domain warping; higher values emphasise finer edges. | | min_thresh | FLOAT | -1.0 | Lower phase threshold for binary edge extraction. | | max_thresh | FLOAT | 0.003 | Upper phase threshold for binary edge extraction. | | morph_flag | BOOLEAN | True | When enabled, applies morphological closing and opening to clean up the binary edge map. When disabled, outputs a soft (normalised) phase image instead. | | output_mode | COMBO | grayscale | grayscale – white edges on black background. rgb_overlay – edges drawn in red over the original image. |

Outputs

| Name | Type | Description | |---|---|---| | edge_image | IMAGE | Processed edge map (binary or overlay depending on output_mode). | | phase_image | IMAGE | Raw PST phase output normalised to 0–1, useful for further processing or visualisation. |

Algorithm

PST processes each image through the following steps:

1. Convert input to grayscale (luminance weights: R×0.2126, G×0.7152, B×0.0722).
2. Compute the 2D FFT and shift to centre zero frequency.
3. Apply a Gaussian low-pass filter in the frequency domain.
4. Apply the PST kernel — a warped phase ramp:

$$K(r) = \frac{W \cdot r}{1 + W \cdot r} \cdot e^{-j , S \cdot r , / , (1 + r)}$$

where $r$ is the radial frequency and $W$, $S$ are the warp and phase strength parameters.

5. Compute the inverse FFT.
6. Extract the phase angle of the complex result.
7. Threshold the phase to produce a binary edge map, then apply morphological clean-up (if morph_flag is enabled).

Parameter Tuning Guide

• More edges / finer detail — increase phase_strength or warp_strength.
• Fewer false edges / less noise — increase lpf_sigma or tighten the threshold range (min_thresh closer to 0, max_thresh closer to 0).
• Soft/continuous edge map — disable morph_flag; use the phase_image output.
• Overlay on original — set output_mode to rgb_overlay.

Example Workflow

Load Image → Phase Stretch Transform → Preview Image
                       ↓
              (edge_image or phase_image)

Screenshots

Requirements

• Python ≥ 3.10
• numpy
• scipy
• torch (provided by ComfyUI)

License

MIT