pypi / kornia
kornia
What changed in kornia 0.8 to 0.9, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-23.
9 breaking1 deprecations19 new APIs and options14 performance6 behavior and defaults23 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initkornia 0.921 changes
Breaking changes 8
Recheck outputs for warps, resize, RANSAC, and integer-image processing because corrected behavior can change results without raising.
Several feature constructors now download pretrained weights by default; select the documented offline or random option when needed.
The documented runtime requirement is Python ≥3.11.
A positive angle turns clockwise, unlike rotate and RandomRotation; account for the different direction when applying transforms.
The Ivy-powered entry point no longer performs conversion; use documented imports such as kornia.image for image helpers.
The Ivy-powered entry point no longer performs conversion; use documented imports such as kornia.image for image helpers.
The Ivy-powered entry point no longer performs conversion; use documented imports such as kornia.image for image helpers.
The documented runtime requirement is torch ≥2.5.1.
New APIs and options 3
An API for repeated feature extraction is available.
A batched matching API is available to use instead of implementing batched matching yourself.
RANSAC(compile=True) is available, but additionally requires torch ≥2.14 and has supported option combinations to check.
Performance 5
Augmentations are reported to perform 10x fewer host synchronizations.
CLAHE is reported as 2000x faster on MPS.
Gaussian blur is reported as 16x faster on Apple CPU.
Morphology operations are reported as 4x faster.
Unsharp mask is reported as 3x faster on Apple CPU.
Fixes and other changes 5
Set KORNIA_RANSAC_AOT=0 as an alternative workaround for the compiled-RANSAC cache crash, at a performance cost.
The documented runtime requirement is kornia_rs ≥0.2.0.
Use documented imports such as kornia.image for image helpers.
Keep learnable bilateral-filter sigmas in float32.
Half-precision warps and geometry are highlighted as working, but support is operation-specific and should be checked before relying on blanket half-precision support.
kornia 0.851 changes
Breaking changes 1
Support for PyTorch versions below 2.0.0 was dropped.
Deprecations 1
Deprecated utilities in kornia.utils begin emitting DeprecationWarning, targeting future removal.
New APIs and options 16
Adds AdaptiveQuadInterp3d and scale-space detector speedups.
Adds the ALIKED local feature detector and descriptor.
Adds half-precision support for float16 and bfloat16.
Adds iterative_quad_interp3d using Cramer's rule.
Adds KORNIA_CHECK_IMAGE_LAYOUT for image layout validation.
Adds mutual information losses and normalized mutual information losses.
Adds Open-CV style thresholding operations to the enhance/threshold module.
Adds fill_value support for RandomAffine.
Adds support for padding_mode="fill".
Adds rectangular and gaussian kernels for mutual information loss.
Adds a closed-form quartic solver.
Adds XFeat and new weights.
Adds an Adaptive Discriminator Augmentation preset.
Adds color space conversion support to Image.
Adds Otsu thresholding for automatic image segmentation.
Implements quaternion averaging.
Performance 9
denormalize_points_with_intrinsics is approximately 2× faster.
depth_to_normals is approximately 3× faster.
Vectorizes a loop to optimize bbox_to_mask3d.
Speeds up compose_transformations by 24%.
Speeds up euclidean_distance by 150% with more efficient tensor operations.
Avoids allocating a 2N x 2N weight matrix.
Speeds up normalize_min_max by 64%.
Speeds up point_line_distance by 28%.
Speeds up relative_transformation by 33%.
Behavior and default changes 6
distance_transform now supports 3D volumes.
mean_iou_bbox now supports xywh and cxcywh bounding box formats.
VisualPrompter now supports batched inference.
The deprecation decorator was removed from depth_to_3d.
Batch functionality was updated.
The resize utility now supports MPS.
Fixes and other changes 18
Fixes BgrToRgba calling the wrong conversion function.
Validates that sigma values are positive.
Fixes device and dtype propagation.
Fixes device and dtype propagation.
Fixes the ONNX io_name_conversion bug and updates variadic type hints.
Fixes MPS-specific runtime errors and skips tests unsupported on MPS.
Adds support for 3D tensors with shape (C, H, W).
Fixes same_on_batch behavior.
Fixes Resize ONNX export compatibility.
Harmonizes dtypes in transform_points batch matrix multiplication.
Fixes handling of a single-pixel instance in the upper-left corner.
Fixes the gradient check for convert_points_from_homogeneous.
Fixes an IndexError in the draw_line method.
Ensures saved grayscale images include a channel dimension (HxWx1).
Fixes a blocking operation in hsv_to_rgb.
Fixes gradient issues in quaternion scalar arithmetic.
Fixes the transformation matrix so it reflects actual image scaling.
Corrects same_on_frame and same_on_batch logic.