Encoders

Encoder choice sets the representation space. In soma, the encoder is selected by EncoderConfig.name and configured through runtime behavior fields such as precision, batch size, and output variant. Geometry is handled through preprocessing, not the encoder config itself.

Encoding methods

• Single encoder — pick one preset from the model zoo below (the common case).

• Composite (multi-encoder) — concatenate several presets into one richer per-position vector: details · tutorial.

The main configuration object is soma.config.EncoderConfig.

class soma.config.EncoderConfig(name, precision=None, batch_size=32, adaptive_batching=False, output_variant=None, allow_non_recommended_settings=False, save_tile_features=False)

Bases: object

Foundation-model encoder selection and model-adjacent settings.

name selects the encoder preset. output_variant exposes preset-specific feature variants when the encoder supports them. allow_non_recommended_settings opts into slide2vec’s warning-only mode when intentionally sweeping non-default runtime settings.

Model Zoo

The tile encoder zoo below is grouped by output dimension for easier scanning, with entries inside each bucket still following the existing date ordering.

Tile-level encoders

Preset	Output dim	Spacing	Notes
lunit	384	0.5	Kang et al. (2023)
prost40m	384	0.5	Grisi et al. (2026)
conch	512	0.5	Lu et al. (2024)
phikon	768	0.5	Filiot et al. (2023)
conchv15	768	0.5	Lu et al. (2024)
hibou-b	768	0.5	Nechaev et al. (2024)
h0-mini	768 / 1536	0.5	Filiot et al. (2024)
phikonv2	1024	0.5	Filiot et al. (2024)
hibou-l	1024	0.5	Nechaev et al. (2024)
uni	1024	0.5	Chen et al. (2024)
musk	1024 / 2048	0.25, 0.5, 1.0	Xiang et al. (2024)
gpfm	1024	0.5	Ma et al. (2024)
mstar	1024	0.5	Wang et al. (2024)
virchow	1280 / 2560	0.5	Vorontsov et al. (2024)
virchow2	1280 / 2560	0.5, 1.0, 2.0	Zimmermann et al. (2024)
uni2	1536	0.5	Chen et al. (2024)
gigapath	1536	0.5	Xu et al. (2024)
h-optimus-0	1536	0.5	Saillard et al. (2024)
h-optimus-1	1536	0.5	Saillard et al. (2024)
midnight	3072	0.25, 0.5, 1.0, 2.0	Karasikov et al. (2025)
genbio-pathfm	4608	0.5	GenBio AI (2024)

Natural-image control

A non-pathology baseline that shares the tile-encoder interface, for measuring how much pathology pretraining actually contributes.

Preset	Output dim	Spacing	Notes
dinov2-vitb14	768	0.5	Oquab et al. (2024)

Slide-level encoders

Preset	Tile encoder	Output dim	Notes
gigapath-slide	gigapath	768	Xu et al. (2024)
titan	conchv15	768	Ding et al. (2024)
prism	virchow (cls_patch_mean)	1280	Shaikovski et al. (2024)
moozy-slide	lunit	768	Kotp et al. (2026)

Patient-level encoders

Preset	Tile encoder	Output dim	Notes
moozy	lunit	768	Kotp et al. (2026)

Compatibility is enforced by the code and by PipelineConfig validation. Use this page to choose a valid starting point, then let the runtime validate the final combination. The spacing table is a reference for selecting preprocessing geometry, not a separate encoder knob.

Discovery helpers

Use soma.list_models() when you want the available encoder presets in code. Pass level="tile", "slide", or "patient" to narrow the list.