Pytorch-Filters

A collection of image edge detection filters built with PyTorch — fully vectorized, GPU-compatible, and easy to drop into any pytorch pipeline.

Filters

• Canny — classic multi-stage edge detector with Gaussian blur, Sobel gradients, non-maximum suppression, and hysteresis
• DoG (Difference of Gaussians) — fast approximation of the Laplacian of Gaussian for edge detection
• XDoG (Extended Difference of Gaussians) — stylized edge detection with soft thresholding, great for non-photorealistic rendering

Installation

pip install pytorch-filters

Quick Start

import torch
import torchvision.transforms.functional as TF
from PIL import Image
from pytorch_filters import canny, difference_of_gaussians, ex_difference_of_gaussians

# Load image as tensor [1, 1, H, W]
img = TF.to_tensor(Image.open("photo.jpg").convert("L")).unsqueeze(0)

edges = canny(img)
dog   = difference_of_gaussians(img, sigma=1.4, k=1.6)
xdog  = ex_difference_of_gaussians(img, sigma=1.0, tau=0.99, phi=100)

Demo

pip install -r requirements.txt
python demo.py images/demo1.jpg

This will display the original image alongside Canny, DoG, and XDoG results side by side.

Requirements

• Python 3.10+
• PyTorch 2.0+
• torchvision
• NumPy