Implementation:Kornia Kornia SSIM Loss

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

Overview

SSIM Loss computes the Structural Dissimilarity (DSSIM) loss based on the Structural Similarity Index Measure (SSIM) for 2D images.

Description

The SSIM Loss computes the Structural Dissimilarity (DSSIM), which is derived from the Structural Similarity Index Measure (SSIM). SSIM is a perceptual metric that quantifies image quality degradation based on luminance, contrast, and structural information.

The DSSIM loss is defined as:

$\text{loss}(x,y)=\frac{1-\text{SSIM}(x,y)}{2}$

The SSIM computation uses a Gaussian kernel of configurable window size to smooth the images before computing the similarity. The loss is clamped to [0, 1].

The implementation supports two padding modes:

• same: Pads the input to produce output of the same spatial size.
• valid: Uses only the valid convolution area, matching the MATLAB reference implementation.

Usage

Import this loss for image generation, super-resolution, and reconstruction tasks where perceptual quality is important. SSIM-based losses capture structural differences that pixel-level losses (L1, L2) may miss, leading to perceptually better results.

Code Reference

Source Location

• Repository: Kornia
• File: kornia/losses/ssim.py
• Lines: 1-129

Signature

def ssim_loss(
    img1: torch.Tensor,
    img2: torch.Tensor,
    window_size: int,
    max_val: float = 1.0,
    eps: float = 1e-12,
    reduction: str = "mean",
    padding: str = "same",
) -> torch.Tensor: ...

class SSIMLoss(nn.Module):
    def __init__(
        self,
        window_size: int,
        max_val: float = 1.0,
        eps: float = 1e-12,
        reduction: str = "mean",
        padding: str = "same",
    ) -> None: ...
    def forward(self, img1: torch.Tensor, img2: torch.Tensor) -> torch.Tensor: ...

Import

from kornia.losses import SSIMLoss
from kornia.losses import ssim_loss

I/O Contract

Inputs

Name	Type	Required	Description
window_size	int	Yes	Size of the Gaussian kernel for smoothing
max_val	float	No	Dynamic range of the images (default: 1.0)
eps	float	No	Small value for numerical stability (default: 1e-12)
reduction	str	No	Reduction mode: 'none', 'mean' (default), or 'sum'
padding	str	No	Padding mode: 'same' (default) or 'valid'
img1	torch.Tensor	Yes	First input image with shape (B, C, H, W)
img2	torch.Tensor	Yes	Second input image with shape (B, C, H, W)

Outputs

Name	Type	Description
loss	torch.Tensor	DSSIM loss value in [0, 1]; scalar for 'mean'/'sum', per-pixel for 'none'

Usage Examples

import torch
from kornia.losses import SSIMLoss

# Create sample image tensors
input1 = torch.rand(1, 4, 5, 5)
input2 = torch.rand(1, 4, 5, 5)

# Using the module API
criterion = SSIMLoss(window_size=5)
loss = criterion(input1, input2)

# With custom parameters
criterion_custom = SSIMLoss(
    window_size=11,
    max_val=1.0,
    eps=1e-12,
    reduction="mean",
    padding="valid",
)
loss = criterion_custom(input1, input2)

Related Pages

• Principle:Kornia_Kornia_Perceptual_Quality_Loss