Implementation:Kornia Kornia Load Image

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

Overview

Concrete tool for loading images from disk and converting to PyTorch tensors provided by the Kornia library.

Description

The load_image function reads image files using the kornia_rs Rust backend for high performance. It supports automatic color space conversion via the ImageLoadType enum, which provides the following modes:

• RGB32 -- 3-channel RGB, float32 normalized to [0, 1]
• GRAY32 -- 1-channel grayscale, float32 normalized to [0, 1]
• RGB8 -- 3-channel RGB, uint8 [0, 255]
• GRAY8 -- 1-channel grayscale, uint8 [0, 255]
• RGBA8 -- 4-channel RGBA, uint8 [0, 255]
• UNCHANGED -- original format preserved

The function returns tensors in (C, H, W) format with float32 [0, 1] normalization for 32-bit modes and uint8 [0, 255] for 8-bit modes. Device placement is handled automatically via the device parameter.

Usage

Import when you need to load images from disk into kornia/PyTorch pipelines. Replaces PIL/OpenCV loading with native tensor output, eliminating manual conversion steps.

Code Reference

Source Location

• Repository: kornia
• File: kornia/io/io.py
• Lines: L83-146

Signature

def load_image(
    path_file: str | Path,
    desired_type: ImageLoadType = ImageLoadType.RGB32,
    device: Union[str, torch.device, None] = "cpu",
) -> torch.Tensor

Import

from kornia.io import load_image, ImageLoadType

I/O Contract

Inputs

Name	Type	Required	Description
path_file	Path	Yes	Path to the image file on disk
desired_type	ImageLoadType	No	Output format and dtype; default RGB32
device	torch.device | None	No	Target compute device; default "cpu"

Outputs

Name	Type	Description
return	torch.Tensor	Image tensor of shape (C, H, W); float32 in [0, 1] for 32-bit types, uint8 in [0, 255] for 8-bit types

Usage Examples

Basic Loading with RGB32

from kornia.io import load_image, ImageLoadType

# Load an image as float32 RGB tensor normalized to [0, 1]
img = load_image("path/to/image.jpg", ImageLoadType.RGB32)
print(img.shape)  # torch.Size([3, H, W])
print(img.dtype)   # torch.float32
print(img.min(), img.max())  # tensor(0.) tensor(1.)

Loading as Grayscale

from kornia.io import load_image, ImageLoadType

# Load as single-channel grayscale
gray = load_image("path/to/image.jpg", ImageLoadType.GRAY32)
print(gray.shape)  # torch.Size([1, H, W])

Loading to GPU

from kornia.io import load_image, ImageLoadType

# Load directly onto GPU
img_gpu = load_image("path/to/image.jpg", ImageLoadType.RGB32, device="cuda")
print(img_gpu.device)  # device(type='cuda', index=0)

Related Pages

Implements Principle

• Principle:Kornia_Kornia_Image_Loading
• Environment:Kornia_Kornia_PyTorch_Python_Environment