Albumentations

AlbumentationsX vs Legacy Albumentations: Practical Benefits

On this page

The legacy albumentations package repository was archived (read-only) on July 10, 2025. No further releases, security patches, bug fixes, pull requests, or issues are possible.

AlbumentationsX continues active development with fifteen releases shipped since the fork and more in progress.

This page covers what changed technically. For the legal side, see the License Guide.

The Legacy Package Is No Longer Maintained

The legacy albumentations package is not "stable" in the sense that it is maintained and reliable. It is frozen. There is a difference:

  • No security patches. If a CVE is filed against a dependency or a vulnerability is found in the library itself, nobody will fix it in the legacy package.
  • No bug fixes. Known issues in replay, serialization, data-loader randomness, and transform behavior were fixed in AlbumentationsX. Those fixes do not exist in the legacy package.
  • No new transforms, no new annotation targets, no performance work. The library stopped evolving in May 2025.

If your team depends on a frozen, unsupported package, that is a risk your security and compliance review should know about.

Which Teams Should Care

Workload or teamWhat AlbumentationsX addsWhy the legacy package becomes limiting
Teams with compliance or security review requirementsOIDC trusted publishing, SBOMs, reproducible lockfiles, SECURITY.mdThe legacy package never shipped supply-chain attestation artifacts and will never receive security patches.
Rotated object detectionNative OBB support across real transform pipelinesYou no longer need custom augmentation geometry for rotated boxes.
Pose estimation or facial landmarksSemantic keypoint label swappingFlips and symmetry transforms can preserve label meaning, not just coordinates.
Medical, remote sensing, hyperspectral, multi-sensor dataStrong multichannel performance work plus non-RGB feature growthThe public benchmark signal is much stronger for multichannel data than for simple RGB-only paths.
Video-heavy preprocessingRepeated video and batch-path optimizations across releasesVideo is being optimized as a primary workload, not only supported incidentally.
Teams with strict debugging and reproducibility needsExact applied-parameter capture and deterministic replayYou can inspect and reconstruct what actually happened to a sample.
Pipelines that carry metadata with imagesuser_data and custom apply_to_X targetsAugmentation can update side-channel data instead of forcing custom wrapper logic around every transform.

Supply-Chain Security and Deployment Reliability

For teams that must audit or attest their dependency chain, these are the changes that matter most:

  • OIDC trusted publishing (no long-lived PyPI API tokens)
  • SBOM generation with SHA-256 checksums for every release asset
  • Reproducible builds via committed lockfile
  • A formal SECURITY.md with disclosure process and SLAs
  • OpenCV as an explicit optional dependency rather than an implicit heuristic-driven install
  • Selectable resize backends (opencv, pillow, pyvips)

Trusted publishing, reproducible lockfiles, and SBOMs are infrastructure that the legacy package never shipped. Teams with compliance or security review requirements get artifacts they can point to.

Capabilities Added After the Fork

Oriented bounding boxes

AlbumentationsX added and expanded native oriented bounding box support in BboxParams, covering pad/crop workflows, a wide transform set, and the cxcywh coordinate format.

Rotated detection is not a corner case in OCR, aerial imagery, industrial inspection, and document analysis. Without native OBB support, teams end up writing custom geometry code around augmentation rather than using augmentation as infrastructure.

Semantic keypoint label swapping

Automatic semantic keypoint label remapping for orientation-changing transforms such as HorizontalFlip, VerticalFlip, D4, and SquareSymmetry.

In pose, face landmarks, and structured keypoint datasets, flipping geometry without flipping semantic labels silently corrupts supervision. AlbumentationsX can keep "left" and "right" consistent inside the augmentation pipeline itself.

Pipeline extensibility

  • user_data target lets a pipeline carry arbitrary Python objects through augmentation and optionally update them per transform. Useful for captions, camera intrinsics, sensor calibration data, and confidence maps.
  • Custom apply_to_X targets let teams make augmentation logic first-class for project-specific targets without forking Compose.
  • Compose arithmetic for adding, subtracting, and prepending transforms programmatically.

Exact replay and augmentation debugging

Exact applied-parameter capture with save_applied_params=True and replay through Compose.from_applied_transforms(...). This records the concrete sampled values that were used for a given sample, not just the transform ranges declared in the pipeline.

This makes augmentation debuggable. You can answer "what exactly happened to this sample?" with concrete sampled values, inspect failed samples, and decide whether a sampled variant is realistic for the deployment domain or whether the augmentation is too aggressive.

3D, volumes, and letterboxing

  • GridShuffle3D for volume, mask3d, and keypoints.
  • LetterBox, the standard scale-to-fit-with-padding preprocessing step used in YOLO-style pipelines, with full target support (images, masks, HBB, OBB, keypoints, volumes).

New transforms

Dithering, AtmosphericFog, ChannelSwap, FilmGrain, Halftone, LensFlare, Vignetting, GridMask, WaterRefraction

The benchmark page itself shows rows where AlbumentationsX has transforms with no legacy package comparison row. That is the clearest sign that the comparison is about feature surface area that continued to grow on one side only.

Performance

All speedup numbers below come from the public comparison at AlbumentationsX vs Legacy Albumentations. The benchmark code and output artifacts are public at albumentations-team/benchmark. Teams can rerun the comparison on their own hardware.

Multichannel

The 9-channel benchmark is the strongest public performance signal. AlbumentationsX wins 62 / 67 head-to-head comparisons with a median speedup midpoint of 1.45x. Headline examples:

If your workload is hyperspectral, medical, remote sensing, or anything with > 3 channels, AlbumentationsX is not just "the same API with updated licensing." Full multichannel comparison table.

RGB geometry and distortion

The speedups are concentrated in transforms that dominate augmentation cost once you move past flips and crops. Headline examples:

Full RGB comparison table.

Video

AlbumentationsX is ahead in 59 / 89 head-to-head video transforms against the legacy package. Headline examples:

Full video comparison table.

Performance optimization is ongoing across releases, with continued work on video and batch paths, C++ kernels for normalization and brightness/contrast, and dropout transform internals.

Get Started

pip install albumentationsx

AlbumentationsX is free under AGPL-3.0 for open-source projects and evaluation. For proprietary software, a commercial license removes the AGPL obligations with unlimited developer seats and a single agreement. See the License Guide for AGPL vs commercial details.

Get a quote

Sources