Anomalib Files

What's Changed

Changed

• 📝 docs: Add Kaputt dataset example notebook and extend datamodule API to support paper's four training settings by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3484

Fixed

• 🐞 model: Fix CFA Descriptor multi-scale fusion to interpolate pooled features instead of raw backbone features by @wsora in https://github.com/open-edge-platform/anomalib/pull/3474
• 🐞 deps: Remove deprecated CUDA 12.4 (cu124) dependency option to resolve Trivy security scan issues by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3476

New Contributors

• @wsora made their first contribution in https://github.com/open-edge-platform/anomalib/pull/3474

Full Changelog: https://github.com/open-edge-platform/anomalib/compare/lib/v2.3.1...lib/v2.3.2

---

API Changes

Removal of cu124 CUDA Extra (breaking changes due to Security Scan Issues)

The cu124 (CUDA 12.4) installation extra has been removed due to Trivy security scan issues. All references have been migrated to cu130 (CUDA 13.0).

[!NOTE] If you still prefer CUDA 12, you can use cu126 instead of cu124

Before (v2.3.1):

:::bash
# uv
uv pip install "anomalib[cu124]"
uv sync --extra cu124

# pip
pip install "anomalib[cu124]"

After (v2.3.2):

:::bash
# uv
uv pip install "anomalib[cu130]"
uv sync --extra cu130

# pip
pip install "anomalib[cu130]"

Note: The cu130 extra now requires torch>=2.9.0, torchvision>=0.24.0, and adds torchcodec>=0.9.0 as a dependency. If you were using cu124 with older torch versions (torch>=2.6.0,<=2.7.1), be aware that the minimum torch version has increased. If you need CUDA 12.x support with older torch versions, use the cu126 extra instead.

Docker users: The Dockerfile now syncs with --extra cu130 instead of --extra cu124.

---

Deprecations (will be removed in v2.6.0)

Kaputt Dataset: use_reference and reference_only Parameters

The use_reference boolean parameter on Kaputt (datamodule) and KaputtDataset has been replaced by the new image_mode enum parameter, which supports three training modes from the Kaputt paper.

Similarly, string values for image_type are deprecated in favor of the ImageType enum.

image_mode replaces use_reference

Before (v2.3.1):

:::python
from anomalib.data import Kaputt

# Default: query images only
datamodule = Kaputt()

# Include reference images
datamodule = Kaputt(use_reference=True)

After (v2.3.2):

:::python
from anomalib.data import Kaputt
from anomalib.data.datasets.image.kaputt import ImageMode

# Default: query images only (unchanged behavior)
datamodule = Kaputt()

# Include reference images
datamodule = Kaputt(image_mode=ImageMode.QUERY_AND_REFERENCE)

# NEW: Reference images only (for memory bank approaches like PatchCore)
datamodule = Kaputt(image_mode=ImageMode.REFERENCE_ONLY)

The three available modes are:

ImageMode Value	Description
QUERY_ONLY	Use only query images (default, same as previous use_reference=False)
QUERY_AND_REFERENCE	Use both query and reference images (same as previous use_reference=True)
REFERENCE_ONLY	Use only reference (defect-free) images for building memory banks

Backward compatibility: Passing use_reference=True/False still works in v2.3.2 but emits a DeprecationWarning. This will be removed in v2.6.0.

ImageType enum replaces string image_type

Before (v2.3.1):

:::python
datamodule = Kaputt(image_type="crop")

After (v2.3.2):

:::python
from anomalib.data.datasets.image.kaputt import ImageType

datamodule = Kaputt(image_type=ImageType.CROP)

Backward compatibility: Passing image_type as a string still works in v2.3.2 but emits a DeprecationWarning. This will be removed in v2.6.0.

Kaputt Dataset: category Parameter

The Kaputt datamodule now accepts a category parameter to filter by material type (e.g., "book_other"). The current default is "all" (loads all categories), which emits a deprecation warning. In v2.6.0, the default will change to "book_other".

:::python
# Suppress the deprecation warning by specifying a category explicitly
datamodule = Kaputt(category="book_other")