AI Potluck
Back to Gap Map Infrastructure / Classic ML & computer vision

Kornia

Kornia
open source / Overall score: 4.0(strong)

Differentiable computer-vision library for PyTorch, with operators for geometry and camera models, feature detection and matching, filtering, color, augmentation and losses that run batched on GPUs inside training graphs. It also ships builders for ported pretrained models such as LoFTR, LightGlue, DISK, SAM, YuNet and RT-DETR, a zoo now frozen for expansion. It is a non-profit project funded by sponsorship.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public(github.com/kornia/kornia)
core features withheld
no — non-profit project funded through Open Collective

Apache-2.0 and built from the public repository, which has no enterprise, commercial or pro directory. Kornia is funded by the people who use it, and it sells nothing.

Adoption

4 high confidence
4.0

Measured on monthly PyPI downloads of the kornia package.

Capability

4 medium confidence
4.0

Kornia is mainly the operations vision models are built and trained from, written to sit inside a network, but it also ships ready pretrained models, including SAM, which segments whatever the user points at in any image with no training. Bundling that kind of model is what places Ultralytics here too. Kornia's models are ports of other projects' networks, and the project has stopped adding them.

  • https://kornia.readthedocs.io/en/latest/ recorded 2026-09-27

    Docs landing: "1,200+ differentiable operators"; "Detect and segment with pretrained models YuNet faces, RT-DETR objects, SAM masks, DexiNed edges - one line to build, weights on first use. 14 model families".

  • https://kornia.readthedocs.io/en/latest/models.html recorded 2026-09-27

    Models page: "Each builder returns a configured model with pretrained weights." and "Pretrained weights are downloaded on first use".

  • https://raw.githubusercontent.com/kornia/kornia/main/README.md recorded 2026-09-27

    README: "The model zoo is currently frozen for expansion while maintainer bandwidth concentrates on the core. Shipped models (LoFTR, LightGlue, DISK, DeDoDe, SAM, and friends) stay available and maintained".

Verified 2026-09-27