English | 简体中文
Docs •
Try it Now •
Tutorials •
Examples •
Blog •
Community
Kornia is a differentiable computer vision library that provides a rich set of differentiable image processing and geometric vision algorithms. Built on top of PyTorch, Kornia integrates seamlessly into existing AI workflows, allowing you to leverage powerful batch transformations, auto-differentiation and GPU acceleration. Whether you're working on image transformations, augmentations, or AI-driven image processing, Kornia equips you with the tools you need to bring your ideas to life.
📢 Direction: Kornia is becoming the reference implementation and executable specification for differentiable computer vision and geometry in the PyTorch ecosystem — explicit conventions, conformance tests, and honest benchmarks over API growth. Read the Roadmap.
When to use Kornia
Kornia pays off where a hand-written version gets long or fragile, or where the alternatives leave the GPU:
- Breadth. 1,200+ public operators, most with no torchvision equivalent: camera models and
epipolar geometry, minimal solvers and RANSAC, feature detection, description and matching, edge detection,
morphology, Lab and Bayer color, image losses and metrics. torchvision overlaps mainly on resizing, cropping,
flipping, color jitter and blur.
- Per-sample parameters in one batched call.
gaussian_blur2d(x, (9, 9), sigma) takes a (B, 2) sigma;
rotate takes a (B,) angle, warp_affine / warp_perspective take (B, 2, 3) / (B, 3, 3) matrices and
adjust_brightness a (B,) factor. torchvision's functional ops apply one parameter per call. From Kornia 0.9.1,
a sigma tensor and the enhancement value checks (gamma, contrast, solarize) are validated on the device, so a
CUDA call, or an MPS call on PyTorch ≥ 2.13, needs no host synchronization for them.
- Batched minimal solvers and robust estimation.
find_fundamental(x1, x2, method="7POINT") solves (B, 7, 2)
samples and returns all real solutions in (B, 3, 3, 3), with zero matrices in the unused slots; the 8-point,
5-point essential (find_essential), DLT homography and PnP (solve_pnp_dlt) solvers are there too. RANSAC
(homography, fundamental and essential models) samples and scores hypotheses on the device; by default its
refinement and final selection run on the host in float64, and compile=True (PyTorch ≥ 2.14, CPU and CUDA) runs
the whole estimation as one compiled graph on the device.
- Differentiable geometry. Depth to 3D and back (
depth_to_3d_v2, unproject_points, project_points),
warp_frame_depth, remap, camera models with Brown–Conrady and Kannala–Brandt distortion, and Lie groups (So3,
Se3) keep gradients with respect to depth, pose and camera parameters.
- Annotated augmentation.
AugmentationSequential applies one sampled transform to images, masks, boxes and
keypoints together and can invert it, in 2D; the 3D augmentations transform volumes, without annotations or an
inverse. On CUDA, augmentations move their sampled parameters to the
GPU without blocking it, and a sync-guard test keeps every public augmentation free of avoidable host
synchronizations.
- Local features on the GPU. SIFT and HardNet, DISK, ALIKED, XFeat, DeDoDe, LoFTR and LightGlue, with
mutual-nearest-neighbour and ratio-test matchers.
These operators run on the input's device with no NumPy round trip, and most are differentiable; robust estimation
(RANSAC) and the extraction methods of learned features such as DeDoDe and XFeat run without gradients. For a
single fixed-parameter operation, such as one blur or one resize, a few lines of PyTorch or torchvision are just as
good.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
