Implementation:Kornia Kornia Gaussian Blur

Knowledge Sources	Kornia
Domains	Vision, Image_Filtering
Last Updated	2026-02-09 15:00 GMT

Overview

Applies Gaussian blur to 2D image tensors by convolving with a Gaussian kernel, with support for batched per-element sigma values and separable filtering.

Description

This module is part of the Kornia library's filters subpackage. It provides the gaussian_blur2d function and GaussianBlur2d nn.Module class for smoothing images using a Gaussian kernel. The Gaussian kernel is parameterized by kernel_size and sigma (standard deviation), and is applied independently to each channel. The implementation supports both separable filtering (default, more efficient -- two sequential 1D convolutions) and non-separable 2D convolution. The sigma parameter can be a tuple (applied uniformly to all batch elements) or a torch.Tensor of shape (B, 2) for per-batch-element sigma values. A deprecated alias gaussian_blur2d_t is also provided.

Usage

Import this function or class whenever you need to apply Gaussian smoothing to images, such as for noise reduction, scale-space construction, or as a preprocessing step before edge detection.

Code Reference

Source Location

• Repository: Kornia
• File: kornia/filters/gaussian.py
• Lines: 1-178

Signature

def gaussian_blur2d(
    input: torch.Tensor,
    kernel_size: tuple[int, int] | int,
    sigma: tuple[float, float] | torch.Tensor,
    border_type: str = "reflect",
    separable: bool = True,
) -> torch.Tensor: ...

class GaussianBlur2d(nn.Module):
    def __init__(self, kernel_size: tuple[int, int] | int,
                 sigma: tuple[float, float] | torch.Tensor,
                 border_type: str = "reflect",
                 separable: bool = True) -> None: ...
    def forward(self, input: torch.Tensor) -> torch.Tensor: ...

Import

from kornia.filters import gaussian_blur2d, GaussianBlur2d

I/O Contract

Inputs

Name	Type	Required	Description
input	torch.Tensor (B, C, H, W)	Yes	The input image tensor.
kernel_size	tuple[int, int] or int	Yes	Size of the Gaussian kernel. Must be positive odd integers.
sigma	tuple[float, float] or torch.Tensor (B, 2)	Yes	Standard deviation of the Gaussian kernel for (height, width). Must be positive.
border_type	str	No (default "reflect")	Padding mode: constant, reflect, replicate, or circular.
separable	bool	No (default True)	If True, uses separable 1D convolutions for efficiency.

Outputs

Name	Type	Description
output	torch.Tensor (B, C, H, W)	The Gaussian-blurred image, same shape as input.

Usage Examples

import torch
from kornia.filters import gaussian_blur2d, GaussianBlur2d

# Functional Gaussian blur
input = torch.rand(2, 4, 5, 5)
output = gaussian_blur2d(input, (3, 3), (1.5, 1.5))
print(output.shape)  # torch.Size([2, 4, 5, 5])

# With per-batch sigma using a tensor
sigma_batch = torch.tensor([[1.5, 1.5], [2.0, 2.0]])
output = gaussian_blur2d(input, (3, 3), sigma_batch)

# Module-based Gaussian blur
gauss = GaussianBlur2d((5, 5), (2.0, 2.0), border_type='reflect')
output = gauss(input)
print(output.shape)  # torch.Size([2, 4, 5, 5])

Related Pages

• Principle:Kornia_Kornia_Gaussian_Smoothing