DINOv3

Foundation models like DINO learn representations of their inputs. That is, they generate very high-dimensional numerical descriptions of what you put into them. The models aren't trained on labelled data, but they're trained on some pretext task like "given this image with a cutout, fill in the cutout" (see Masked Auto-Encoders). So the basic output from a model is a vector - often called an embedding. Literally a 1D list of numbers, O(1k)-dimensional. Your goal is to get an embedding that assigns (well) linearly separable vectors for all the things you want to classify.

Vision transformers also output patch tokens, which can be assembled into a low-resolution feature map (w/32, h/32 is common). So what you do with that data depends on the task. Classification can be as simple as linearly classifying the (whole image) embedding. A semantic segmentation task can do the same, but for every pixel. This is why the DINO authors show a PCA representation of a bunch of images, which show that semantically similar objects are grouped together by colour. Object detectors are more complicated, but the key thing is that once you have these pixel-level features, you can use them as input into existing architectures.

Now to your question: face recognition is a specific application of object re-identification (keyword: Re-ID). The way most of these models work is from the whole-image embedding. Normally you'd run a detector to extract the face region, then compute the embedding, put it in a vector database and then query for nearest neighbours using something like the cosine distance. I've only worked in this space for animals, but humans are far more studied. Whether DINOv3 is good enough out-of-the-box I don't know, but certainly there's a lot of literature looking at these sorts of models for Re-ID.

The challenge with Re-ID is that the model has to be able to produce features which discriminate individuals rather than similar looking individuals. For example with the vanilla model, you probably have a very good tool for visual search. But that's not the same task, because if you give it a picture of someone in a field, you'll get back pictures of other people in fields. That usually requires re-training on labelled imagery where you have a few examples of each person. The short answer is that there are already very good models for doing this, and they don't necessarily even need ML to do a decent job (though it might be used for keypoint detection for facial landmarks).