Kornia is a geometric computer vision library for PyTorch that makes classical vision operations — image transformations, camera geometry, feature detection — differentiable and GPU-accelerated, so they can live inside neural network training instead of outside it. It has earned about 11,400 stars on GitHub and is in active development in 2026.
Traditional vision pipelines use libraries like OpenCV, which are fast but opaque to gradient descent: you cannot backpropagate through them, batch them on a GPU, or make them part of a learned model. Kornia reimplements the core operations as native PyTorch tensor code. Its modules cover color space conversion and image enhancement, geometric transformations and augmentation, filtering, local feature detection and description with matching, camera models and epipolar geometry, and losses designed for geometric tasks. Because everything is differentiable, you can train a network to predict the parameters of a classical pipeline, augment data on the GPU in the same graph as training, or build spatial-AI systems — robotics, SLAM, 3D reconstruction, autonomous driving perception — where geometry and learning must interact. Kornia is maintained by an open-source community with roots in the research groups that popularized differentiable vision, and it serves as the geometry layer under many PyTorch-based vision projects.
Kornia is completely free and open source under the Apache-2.0 license. There are no paid tiers; it runs wherever PyTorch runs.
Kornia is for computer vision researchers and engineers who need classical geometry and image processing inside deep learning systems — training with on-GPU augmentation, building spatial-AI and robotics perception, or learning the parameters of vision pipelines end to end.