Implementation:Kornia Kornia Tensor To Image

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

Overview

Concrete tool for converting PyTorch tensors to numpy image arrays provided by Kornia.

Description

The tensor_to_image function converts torch.Tensor images from (C, H, W) or (B, C, H, W) format to numpy.ndarray in (H, W, C) or (H, W) format suitable for matplotlib or PIL display.

Key features:

• Automatic channel permutation from CHW to HWC format
• CPU transfer handled internally (works with GPU tensors)
• Grayscale handling -- single-channel tensors produce (H, W) arrays
• Optional contiguous memory layout via force_contiguous for compatibility with libraries requiring contiguous arrays
• Batch dimension handling via keepdim parameter

Usage

Import when you need to visualize tensor outputs using matplotlib, PIL, or save to disk. This is the standard way to move from Kornia's tensor format back to the numpy/image ecosystem.

Code Reference

Source Location

• Repository: kornia
• File: kornia/image/utils.py
• Lines: L167-225

Signature

def tensor_to_image(
    tensor: torch.Tensor,
    keepdim: bool = False,
    force_contiguous: bool = False,
) -> np.ndarray

Import

from kornia.utils import tensor_to_image
# or
from kornia import tensor_to_image

I/O Contract

Inputs

Name	Type	Required	Description
tensor	torch.Tensor	Yes	Image tensor in (C, H, W) or (B, C, H, W) format
keepdim	bool	No	Preserve batch dimension in output; default False
force_contiguous	bool	No	Ensure output array has contiguous memory layout; default False

Outputs

Name	Type	Description
return	numpy.ndarray	Image array in (H, W, C) format for color images, (H, W) for grayscale

Usage Examples

Basic Conversion for Matplotlib Display

import torch
import matplotlib.pyplot as plt
from kornia.utils import tensor_to_image

# Assume img_tensor is a (3, H, W) float32 tensor from a Kornia pipeline
img_tensor = torch.randn(3, 256, 256).clamp(0, 1)

# Convert to numpy for matplotlib
img_np = tensor_to_image(img_tensor)
print(img_np.shape)  # (256, 256, 3)

plt.imshow(img_np)
plt.title("Converted Image")
plt.show()

Batch Tensor Conversion

import torch
from kornia.utils import tensor_to_image

# Batch of 4 images
batch_tensor = torch.randn(4, 3, 128, 128).clamp(0, 1)

# Convert each image in the batch
for i in range(batch_tensor.shape[0]):
    img_np = tensor_to_image(batch_tensor[i])
    print(f"Image {i} shape: {img_np.shape}")  # (128, 128, 3)

Grayscale Conversion

import torch
from kornia.utils import tensor_to_image

# Single-channel grayscale tensor
gray_tensor = torch.randn(1, 256, 256).clamp(0, 1)

img_np = tensor_to_image(gray_tensor)
print(img_np.shape)  # (256, 256) -- channel dimension is squeezed

Related Pages

Implements Principle

• Principle:Kornia_Kornia_Tensor_Image_Conversion
• Environment:Kornia_Kornia_PyTorch_Python_Environment