Implementation:Kornia Kornia SmallSR

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

Overview

SmallSR implements a compact super-resolution neural network using efficient sub-pixel convolution for real-time image upscaling, based on the paper Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network by Shi et al.

Description

This module provides two classes within the Kornia library: SmallSRNet, the core super-resolution network that operates on single-channel (luminance) images, and SmallSRNetWrapper, which wraps the network with RGB-to-YCbCr and YCbCr-to-RGB color space conversions for end-to-end processing of color images. The network architecture consists of three convolutional layers with ReLU activations followed by a sub-pixel shuffle layer (nn.PixelShuffle). Pre-trained weights are available and downloaded automatically via CachedDownloader. The module also provides a weight_init function using orthogonal initialization.

Usage

Import this module when you need a lightweight, real-time super-resolution model. Use SmallSRNet directly for grayscale/luminance upscaling, or SmallSRNetWrapper for seamless RGB image super-resolution.

Code Reference

Source Location

• Repository: Kornia
• File: kornia/models/small_sr.py
• Lines: 1-99

Signature

class SmallSRNet(nn.Module):
    def __init__(self, upscale_factor: int, inplace: bool = False, pretrained: bool = True) -> None: ...
    def load_from_file(self, path_file: str) -> None: ...
    def forward(self, x: torch.Tensor) -> torch.Tensor: ...

class SmallSRNetWrapper(nn.Module):
    def __init__(self, upscale_factor: int = 3, pretrained: bool = True) -> None: ...
    def forward(self, input: torch.Tensor) -> torch.Tensor: ...

def weight_init(model: nn.Module) -> None: ...

Import

from kornia.models.small_sr import SmallSRNet, SmallSRNetWrapper

I/O Contract

Inputs (SmallSRNet)

Name	Type	Required	Description
upscale_factor	int	Yes	Factor by which to increase image resolution.
inplace	bool	No	Whether ReLU should operate in-place (default False).
pretrained	bool	No	Whether to load pre-trained weights (default True).

Inputs (SmallSRNetWrapper)

Name	Type	Required	Description
upscale_factor	int	No	Factor by which to increase image resolution (default 3).
pretrained	bool	No	Whether to load pre-trained weights (default True).

Outputs

Name	Type	Description
output (SmallSRNet)	torch.Tensor	Upscaled single-channel image of shape (B, 1, H*factor, W*factor).
output (SmallSRNetWrapper)	torch.Tensor	Upscaled RGB image of shape (B, 3, H*factor, W*factor).

Architecture

The SmallSRNet consists of:

1. conv1: Conv2d(1, 64, 5x5) with ReLU
2. conv2: Conv2d(64, 64, 3x3) with ReLU
3. conv3: Conv2d(64, 32, 3x3) with ReLU
4. conv4: Conv2d(32, upscale_factor^2, 3x3)
5. pixel_shuffle: PixelShuffle(upscale_factor)

The SmallSRNetWrapper processes RGB images by:

1. Converting RGB to YCbCr
2. Applying SmallSRNet to the Y (luminance) channel
3. Bicubic-interpolating the Cb and Cr (chrominance) channels
4. Concatenating and converting back to RGB

Usage Examples

import torch
from kornia.models.small_sr import SmallSRNet, SmallSRNetWrapper

# Single-channel super-resolution
sr_net = SmallSRNet(upscale_factor=3, pretrained=True)
y_channel = torch.rand(1, 1, 64, 64)
y_upscaled = sr_net(y_channel)  # shape: (1, 1, 192, 192)

# Full RGB super-resolution
sr_wrapper = SmallSRNetWrapper(upscale_factor=3)
rgb_image = torch.rand(1, 3, 64, 64)
rgb_upscaled = sr_wrapper(rgb_image)  # shape: (1, 3, 192, 192)

Related Pages

• Principle:Kornia_Kornia_Small_Super_Resolution