LazySlide models

The model zoo for LazySlide, the accessible whole slide image analysis framework

lazyslide-models contains all the models consumed by LazySlide. LazySlide itself contains no model code, it is an inference and orchestration layer that discovers models through the MODEL_REGISTRY. Any registered model can be used by name with zero configuration.

Model types

The zoo spans different task types: vision encoders, multimodal (image-text) models, segmentation, slide encoders, tile prediction, classical hand-crafted features, feature prediction, style transfer (HE to spatial omics or IHC stainings), image generation and chat models. Every timm model is also available for feature extraction (e.g. resnet50).

The complete per-task model list, with licenses, gated-access badges and citation keys, is maintained in the Model Zoo Website.

Usage

from lazyslide_models import MODEL_REGISTRY, list_models

list_models("segmentation")          # ['cellpose', 'classpose', 'sam', ...]
model = MODEL_REGISTRY["instanseg"]()  # instantiate (weights download on first use)

In LazySlide, pass the registered name to any function:

zs.tl.feature_extraction(wsi, model="conch")  # any vision or multimodal key
zs.seg.cells(wsi, model="instanseg")

Exported models

Models that cannot be loaded directly from Hugging Face are exported to RendeiroLab/LazySlide-models (permissive and non-commercial licenses) and RendeiroLab/LazySlide-models-gpl (GPL). Reproduction scripts: scripts/export_models.

Contributing a new model

1. Open an issue with the [New Model] label to confirm the model fits the zoo.
2. Follow the integration guide.
3. Open a pull request against main. The CI will test the models in your PR.

Licenses

Some models are non-commercial or gated, and licenses vary. Check each model's registry entry before use, and please cite the original paper.