What the registers fixed

The practical complaint about DINOv2 for dense tasks was the artefacts: a handful of high-norm tokens in low-information patches that behave like scratch space and wreck attention maps if you visualise them. Registers — extra learnable tokens whose only job is to absorb that traffic — clean it up, and the dense feature maps stop having bright holes in the sky.

For my purposes the interesting part is that this is a representation fix, not a loss fix. The model was already learning the right thing; it lacked somewhere to put the bookkeeping.

Used frozen in HNSD-VLM, it behaves the way you want a backbone to behave: you can attach a small trainable head and the features do not fight you. The ablation across backbones is in the plate, embargoed until publication, so I will only say that the choice of encoder mattered more than the fusion design sitting on top of it, which was not what I expected going in.

Caveat worth stating: everything above is about frozen-feature quality on natural images. Fundus photographs are not natural images, and the transfer story there is its own experiment.