Albumentations to Kornia Transform Mapping

On this page

• How to Read This Page
• Migration Rules
• Color and Pixel Transforms
• Illumination, Weather, and Procedural Effects
• Geometric Transforms
• Dropout and Annotation-Aware Transforms
• Channel Notes
• Migration Example
• Related Pages

This page maps Albumentations transforms to the closest Kornia API when Kornia has a practical direct equivalent. Use it as a migration and capability guide, not as a performance page.

How to Read This Page

• A named Kornia transform means Kornia has a built-in API for the same broad operation.
• - means Kornia does not support that transform as a built-in augmentation primitive.
• The table shows built-in Kornia APIs. Custom PyTorch code is not counted as Kornia support.
• Unsupported rows stay visible because transform coverage is part of the library choice.

Migration Rules

• Albumentations expects NumPy arrays, usually per sample in OpenCV-style H,W,C channel-last layout.
• Kornia expects PyTorch tensors, usually batched as B,C,H,W.
• Kornia often exposes same_on_batch, p_batch, align_corners, keepdim, and tensor interpolation settings.
• Albumentations exposes target-aware behavior through A.Compose.
• Kornia AugmentationSequential can coordinate configured data keys such as images, masks, boxes, keypoints, classes, and labels; the tensor/data-key setup is part of the migration.
• Kornia mix augmentations such as MixUp/CutMix work with image/input keys only. They are useful batch-level tensor policies, not target-aware detection or segmentation policy replacements.
• Albumentations Mosaic and CopyAndPaste are per-sample multi-image policies with target handling. Kornia mix/transplant APIs are tensor/batch policies with different target semantics.
• RGB/color-space transforms are not arbitrary-channel transforms in either library.

Color and Pixel Transforms

Illumination, Weather, and Procedural Effects

Geometric Transforms

Dropout and Annotation-Aware Transforms

Channel Notes

Kornia and Albumentations can both operate on non-RGB tensors/arrays for many channel-agnostic transforms. The trap is color semantics: RGBShift, HueSaturationValue, ColorJitter, PlanckianJitter, RandomRain, and RandomSnow are RGB/color transforms.

Blur, crop, resize, flip, affine, perspective, noise, dropout, and normalization are usually safer for arbitrary-channel data. For medical, hyperspectral, remote-sensing, or sensor-fusion inputs, document whether each channel is exchangeable, geometric-only, or color-like before choosing transforms.

Migration Example

Kornia pipelines are tensor-native:

import kornia.augmentation as K

pipeline = K.AugmentationSequential(
    K.RandomHorizontalFlip(p=0.5),
    K.RandomRotation(degrees=15.0, p=0.5),
    K.RandomGaussianNoise(mean=0.0, std=0.1, p=0.2),
    data_keys=["input"],
)

The equivalent Albumentations policy is usually applied per sample before tensor conversion:

import albumentations as A

pipeline = A.Compose(
    [
        A.HorizontalFlip(p=0.5),
        A.Rotate(angle_range=(-15, 15), p=0.5),
        A.GaussNoise(std_range=(0.05, 0.2), p=0.2),
    ],
)

For annotation-aware pipelines, Kornia uses AugmentationSequential data keys:

import kornia.augmentation as K

pipeline = K.AugmentationSequential(
    K.RandomCrop(size=(512, 512), p=1.0),
    K.RandomHorizontalFlip(p=0.5),
    data_keys=["input", "bbox"],
)

Albumentations keeps bbox format and label fields in the augmentation policy:

import albumentations as A

pipeline = A.Compose(
    [
        A.RandomCrop(height=512, width=512, p=1.0),
        A.HorizontalFlip(p=0.5),
    ],
    bbox_params=A.BboxParams(coord_format="pascal_voc", label_fields=["labels"]),
)

Related Pages

• Kornia comparison