TECHNOLOGY PROFILE · TECH-0128
Kornia
Differentiable computer-vision operations for PyTorch.
Reference link only — not reviewed this edition. This profile preserves the prepared catalog’s original descriptions. Advantages depend on the workload; comparisons are editorial, not benchmark results.
Where it has an edge
GPU tensor and differentiable transforms can stay inside a model graph instead of CPU preprocessing.
Best-fit problems
Differentiable augmentation, geometry and vision research.
Tradeoffs & cautions
GPU overhead can exceed benefits for small batches; gradient support varies.
What to test before choosing
Inspect transformed examples; compare task metrics by slice and test annotation integrity and deployment support.
Runtime & hardware
Local CPU/GPU
Access & licensing
Open-source cores; model/add-on licenses vary
Execution
Local / self-hosted
Lifecycle
Not independently audited
Comparison basis
Editorial problem-fit judgement; no head-to-head benchmark performed
Connected concepts
Alternatives to investigate
OpenCV · Albumentations / AlbumentationsX