Kornia Cheat Sheet

Kornia cheat sheet

Kornia is a differentiable computer-vision library built on PyTorch: classic CV ops (filtering, color, geometry, features) that run on the GPU, work in batches, and are differentiable — so they slot into a training graph and you can backprop through them. Think "OpenCV/skimage as tensor ops". Everything operates on (B, C, H, W) float tensors in [0,1]. Highlights: GPU augmentation that keeps images/masks/boxes/keypoints in sync, and deep local-feature matching (LoFTR). Targets Kornia 0.8.

tensors & color augmentation filters & enhance geometry features & losses gotcha ★ most common

Verified 2026-08-24 against the official docs at kornia.org / kornia.readthedocs.io and the kornia/kornia repo (0.8.x). Pure PyTorch — ops accept & return tensors, run on any device, and are autograd-compatible.

Outline

The whole library is tensor-in / tensor-out. The two things that make Kornia special: differentiable geometry and GPU augmentation that transforms labels alongside the image.

9Features & matchingkornia.feature

matcher = K.feature.LoFTR(pretrained="outdoor") correspondences = matcher({"image0": g0, "image1": g1})★
LoFTR detector-free deep matching — dense correspondences between two images, pretrained.
K.feature.DISK.from_pretrained(...) · K.feature.SIFTFeature() · KeyNet...
Learned (DISK) & classic (SIFT/Harris) local features — keypoints + descriptors as tensors.
K.feature.match_snn(desc1, desc2) · LocalFeatureMatcher(...)
Descriptor matching (mutual NN / ratio test) and end-to-end matcher modules.
K.geometry.ransac.RANSAC("homography")(pts1, pts2)
Robustly fit a homography/fundamental matrix from noisy matches.

10Losses & modelskornia.losses

K.losses.ssim_loss(pred, tgt, 5) · psnr_loss(...) · total_variation(...)★
Image-quality losses. Also focal_loss, dice_loss, tversky_loss for segmentation.
K.filters.get_gaussian_kernel2d(...) # build kernels as tensors
Kernel factories when you need a custom differentiable filter.
K.contrib.FaceDetector() · kornia model zoo (segmentation, matching)
Pretrained task models exposed as modules.

!Common gotchasread before shipping

(B, C, H, W) float [0,1]
The universal input format. Passing HWC, uint8, or an unbatched tensor is the #1 error — use image_to_tensor and add the batch dim.
augment on the GPU, in the loop
Kornia augmentation shines when the batch is already on the GPU — not as a per-sample CPU Dataset transform (that throws away the speedup).
use AugmentationSequential for labels
Applying separate random augs to image and mask desynchronizes them. AugmentationSequential with data_keys keeps geometry consistent.
it's differentiable — mind the graph
Ops keep gradients. Wrap in torch.no_grad() for pure preprocessing, or you'll accumulate an unwanted graph/memory.

Worth memorizing

tensor in, tensor out(B, C, H, W) float [0,1] — OpenCV/skimage as PyTorch ops

image_to_tensor / tensor_to_imageHWC uint8 ↔ CHW float bridge

augment on GPU in the loopkornia.augmentation modules, batched & random-per-sample

AugmentationSequential(data_keys=...)image + mask + bbox + keypoints transformed together

warps are differentiableoptimize a homography with backward() — image registration by SGD

get_perspective_transform + warp_perspectivefour point pairs → warp

LoFTR for deep matchingpretrained detector-free correspondences

ssim/dice/focal lossesready-made differentiable image & segmentation losses

no_grad for pure preprocessingops keep gradients by default

Kornia cheat sheet

Outline

Basics

Augment

Process

Geometry

Deep CV

Basics

Augmentation

Process

Geometry

Deep CV

Worth memorizing