Composite encoders

A composite encoder concatenates several frozen encoders’ dense features for segmentation or detection. Each member is cached independently; CompositeDenseFeatureStore combines them at load time. Name the composite encoder in place of a single encoder.

features:
  encoder:
    name: composite
    params:
      members:
        - { name: uni, dense: { feature_kind: cls_attention } }
        - { name: phikon, dense: { feature_kind: cls_attention } }
        - { name: lunit, dense: { feature_kind: patch_features }, member_norm: l2 }

Each member can select its feature kind, attention settings and extraction window (dense, fields of slide2vec’s DenseOptions), its encoder params (params, such as output_variant) and its execution precision and batch size. If omitted, feature_kind resolves to patch_features. All members share the run’s read spacing and supervision size (features.dense); per-member spacing is not supported. Set window_size and overlap in a member’s dense to override the shared settings, for example when an encoder requires a fixed input size.

Concatenation resolution

concat_resolution selects where the members’ grids are combined:

• grid (the default) resamples each member to a common token grid before channels are stacked, which is what a decoder consumes. concat_grid_size sets that grid’s (height, width); by default soma takes the largest height and width across members. This mode treats each grid as spanning the target field of view and ignores individual members’ padding fractions.

• target interpolates each member to its padded encoded size and crops it to the shared supervision target. Channels then stack into (Σd_i, H, W). This preserves each member’s padding and crop geometry even when token-grid sizes differ.

Normalization follows resampling and precedes concatenation. Per-member member_norm accepts none, l2, or layernorm and operates across channels at each position. Defaults are l2 for patch features and none for attention maps. Use it to control differences in feature magnitude across encoders.

References

• Ramchandani et al., Benchmarking Computational Pathology Foundation Models for Semantic Segmentation (2026), arXiv:2602.18747.