Anomalib Files

What's Changed

Added

• 🚀 model: Migrate all DINOv2-based models to timm feature extraction with DINOv3 backbone support by @alexriedel1 in https://github.com/open-edge-platform/anomalib/pull/3627
• 🚀 model: Add output_fmt="NLC" mode to TimmFeatureExtractor for transformer backbones via timm's forward_intermediates API by @alexriedel1 in https://github.com/open-edge-platform/anomalib/pull/3627
• 🚀 model: INP-Former half-precision training support by @alexriedel1 in https://github.com/open-edge-platform/anomalib/pull/3625
• 🚀 docs: Improve PatchCore backbone layer documentation by @rashidrao-pk in https://github.com/open-edge-platform/anomalib/pull/3630

Removed

• ⚠️ model: Remove custom DinoVisionTransformer, DinoV2Loader, and Dinomaly's internal ViT implementation — replaced by timm-based extraction by @alexriedel1 in https://github.com/open-edge-platform/anomalib/pull/3627
• ⚠️ deps: Remove cu118 install extra by @AlexanderBarabanov in https://github.com/open-edge-platform/anomalib/pull/3652

Changed

• 🔧 deps: Replace opencv-python with opencv-python-headless to support headless environments by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3675
• 🔧 deps: Bump minimum torch and torchvision versions for the xpu install extra by @AlexanderBarabanov in https://github.com/open-edge-platform/anomalib/pull/3652
• 🔧 deps: Remove upper bound on torch for cpu, cu126, and rocm install extras to unblock Python 3.14 wheel resolution by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3554
• 🔧 model: AnomalyDINO default encoder_name changed from dinov2_vit_small_14 to vit_small_patch14_dinov2 (timm naming) by @alexriedel1 in https://github.com/open-edge-platform/anomalib/pull/3627
• 🔧 model: Dinomaly, INP-Former, L2BT, PatchFlow encoder parameters updated to use timm backbone names by @alexriedel1 in https://github.com/open-edge-platform/anomalib/pull/3627

Deprecated

• ⚠️ model: anomalib.models.components.dinov2 package emits FutureWarning, use TimmFeatureExtractor instead by @alexriedel1 in https://github.com/open-edge-platform/anomalib/pull/3627
• ⚠️ export: Deprecate the legacy ONNX exporter path (dynamo=False) ahead of anomalib 2.7.0, where the minimum required PyTorch version will increase to 2.10 by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3554

Fixed

• 🐞 data: Update stale MVTec AD/3D download URLs and switch ShanghaiTech to manual-download (dead mirror) with user-facing instructions by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3673
• 🐞 export: Add dynamic_shapes support for torch dynamo ONNX/OpenVINO export while keeping the legacy exporter path available during the 2.7.0 transition by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3554
• 🐞 data: Normalize Split enum to string value before DataFrame comparison in dataset loaders for pandas 3.x compatibility by @won-seoop in https://github.com/open-edge-platform/anomalib/pull/3634 and https://github.com/open-edge-platform/anomalib/pull/3659
• 🐞 data: Fix pandas 3 compatibility by @won-seoop in https://github.com/open-edge-platform/anomalib/pull/3668
• 🐞 typing: Pin mypy python_version to 3.10 to fix spurious dataclass __replace__ incompatibility errors under Python 3.13 by @ashwinvaidya17 in https://github.com/open-edge-platform/anomalib/pull/3674
• 🐞 model: Remove CPU-GPU synchronization in coreset selection loop by @andersendsa in https://github.com/open-edge-platform/anomalib/pull/3651

New Contributors

• @jpggvilaca made their first contribution in https://github.com/open-edge-platform/anomalib/pull/3643
• @dabhi365 made their first contribution in https://github.com/open-edge-platform/anomalib/pull/3614
• @andersendsa made their first contribution in https://github.com/open-edge-platform/anomalib/pull/3651
• @rashidrao-pk made their first contribution in https://github.com/open-edge-platform/anomalib/pull/3630

Full Changelog: https://github.com/open-edge-platform/anomalib/compare/lib/v2.5.0...lib/v2.5.1

API Changes

All DINOv2-based models (AnomalyDINO, Dinomaly, INP-Former, L2BT, PatchFlow) now use TimmFeatureExtractor instead of anomalib's custom DinoVisionTransformer and DinoV2Loader implementations. This change:

• Enables DINOv3 drop-in support — use any timm DINOv3 backbone name (e.g. vit_large_patch16_dinov3.lvd1689m).
• Reduces maintenance surface — removes ~800 lines of custom ViT code.
• Unifies the feature extraction API across CNN and transformer backbones.

See PR [#3627] for full details and benchmark comparisons.

Usage example (DINOv3 with Dinomaly):

:::python
from anomalib.data import MVTecAD
from anomalib.models import Dinomaly
from anomalib.engine import Engine

model = Dinomaly(encoder_name="vit_large_patch16_dinov3.lvd1689m")
engine = Engine(max_epochs=15)
engine.fit(datamodule=MVTecAD(), model=model)