Add a natural-image DINOv2 ViT-B/14 control tile encoder

[clemsgrs]

What

Registers dinov2-vitb14: the original DINOv2 ViT-B/14 self-supervised on LVD-142M natural images (Oquab et al., 2024), shipped by timm as vit_base_patch14_dinov2.lvd142m. Weights are hosted publicly on Hugging Face under timm/vit_base_patch14_dinov2.lvd142m — no gated access, no HF token.

Implements clemsgrs/soma#232.

Why

soma's detection benchmark (Paper 1) needs a non-pathology ViT control to answer "does pathology-pretraining actually pay off for detection?" (and Paper 2 the "catches up late" curve). Almost every pathology tile encoder here (UNI, Virchow, GigaPath, H-optimus, Midnight, …) is a DINOv2-family ViT, so pairing them with a DINOv2 ViT trained on natural images holds architecture and the SSL objective fixed and varies only the pretraining domain — exactly the control the benchmark wants.

How

Structurally a plain TimmTileEncoder, mirroring lunit / prost40m / uni: the dense (encode_tiles_dense) and attention (encode_tiles_attention) paths are inherited unchanged from the timm ViT base, so the control is dense-extraction- and attention-capable exactly like the pathology encoders. dynamic_img_size=True lets the natively-518px backbone run at the 224px detection tile geometry via positional-embedding interpolation (a no-op at native size, verified in the shared dense-extraction suite).

Metadata

field	value	note
level	tile
output dim	768	ViT-B, single default variant
input_size	224	matches every pathology tile encoder → same tile geometry
patch_size	14	224 / 14 = 16 → 16×16 dense grid
precision	fp16
supported_spacing_um	None	spacing-agnostic — no validation constraint (see below)
default_spacing_um	0.5	tiling default when selected by name
source	timm/vit_base_patch14_dinov2.lvd142m	public

Spacing model — split constraint from default

A natural-image model has no intrinsic micron-per-pixel spacing, so advertising 0.5 as a supported/recommended spacing (the way the pathology encoders do) would be a lie: it would make validate_encoder_config reject any other spacing as "non-recommended" when in fact no spacing is more correct than another. This PR splits the previously overloaded supported_spacing_um into its two real roles:

• supported_spacing_um — the validation constraint: the spacing(s) a model was trained/validated for. Off-list requests raise unless allow_non_recommended_settings=True. Set to None here to mark the encoder spacing-agnostic — the spacing check is skipped entirely.
• default_spacing_um — the single spacing a slide is tiled at when the encoder is selected by name with no explicit requested_spacing_um. Set to 0.5 here so the control lands on the pathology encoders' tile geometry and drops in with zero config.

Consequences:

• Selecting dinov2-vitb14 by name still tiles at 0.5 µm/px → matched geometry, zero config (acceptance criterion satisfied).
• Sweeping other task-spacings (e.g. the 0.25 used in the OCELOT campaign) now needs no escape hatch — any requested spacing is accepted as-is, because the model has no spacing to violate.

default_spacing_um is optional and derived from supported_spacing_um when that is a single value, so the ~13 single-spacing pathology encoders need no edits. The multi-spacing guardrail is preserved: virchow2 / musk / midnight (a list of spacings) still require an explicit requested_spacing_um unless they opt in with their own default_spacing_um. Default resolution is kept lazy so tile-size validation never trips over a list/agnostic encoder.

Other judgment calls

• Plain DINOv2 (no registers). Picked vit_base_patch14_dinov2 over the reg4 variant for the cleanest "vanilla ViT" control: a single CLS prefix token, which is the simplest dense/attention layout (like uni/midnight/phikon). The reg4 variant is a trivial swap if you'd prefer to match the register-carrying pathology models (UNI2, H-optimus, Virchow2).
• ViT-B/14 as the size: the mid-size that most directly parallels the ViT-L/g pathology backbones while staying cheap.
• Aliases dinov2 / dinov2-base / dinov2-vitb → dinov2-vitb14.

Tests

New tests/test_dinov2_natimage.py mirrors the existing encoder suite:

• registration/metadata contract (supported_spacing_um is None, default_spacing_um == 0.5) + alias resolution,
• spacing model: name-only selection resolves the 0.5 tiling default, and validate_encoder_config accepts any requested spacing (0.25 / 2.0) with no escape hatch,
• dense grid parity vs timm's get_intermediate_layers oracle at the 224px detection geometry (offline, pretrained=False) — pins spatial registration, not just shape,
• CLS-only per-head attention shape.

The registry split (supported_spacing_um accepts None; new default_spacing_um) lives in slide2vec/encoders/registry.py; the existing multi-spacing guard tests (virchow2/musk/midnight in test_regression_core / test_regression_models) are unchanged and still pass. Also updated the shared exact-membership lists (test_encoder_registry, test_patch_size_metadata, test_regression_core), docs (docs/models.rst), and the README count.

Full suite (-o addopts=""): 483 passed, 4 skipped (the 4 skips are the weight-dependent heavy tests, which run on the nightly heavy workflow). Separately verified end-to-end with the real downloaded weights on CPU: dense (1, 768, 16, 16) · attn (1, 12, 16, 16) · pooled (1, 768) at 224px.

Coverage note: this repo's --cov addopts hit a sqlite lock on my (network) filesystem, so I ran pytest with -o addopts="". Heavy weight-parity test (test_static_patch_size_matches_runtime) is excluded from the PR suite by design and runs on the nightly heavy workflow.