Invisible Model Identifies Facial Features express
Trained on ordinary images, the DINOv3 model unexpectedly began to see faces as a spectrum: on its own, without hints, it isolates the eye, nose, and mouth regions.
Across 200 videos, feature labeling accuracy reached 83% (random guessing is 23%). For tracking lips and eyebrows, it hit 95.5%. But the sharpest 'spectral' separation lies not in the final layer, but in the 18th of 24 layers: there, differences are fivefold, fading by the output.
Compared to CLIP, which captures the overall silhouette, DINOv3 is 16.8 percentage points more accurate at distinguishing fine features. Its architecture, akin to standard visual systems, spontaneously generated a facial map deep within — a veritable spectroscope for features. The frozen representations from the middle of the network are a ready-made tool for decoding facial expressions without any further training.
Original abstract
as written by the authors arXiv:2607.14423
Frozen self-supervised vision models can align parts of generic objects, but it remains unclear whether this correspondence extends to human faces, where global layout is shared while identity-specific appearance varies sharply. We test whether frozen DINOv3 features define a region-level facial coordinate system: a feature space in which eyes, brows, nose, mouth, skin, and hair remain distinguishable across people and across time without face-specific training. Using DINOv3 ViT-L/16 patch embeddings and FaRL only as a face-part labeling interface, we evaluate cross-identity nearest-neighbor matching and temporal label propagation on 200 CelebDF-v2 real videos. DINOv3 achieves 83.0% region-level semantic accuracy under unconstrained cross-identity matching, compared with a 23.0% area-weighted random baseline, and 95.5% temporal tracking accuracy without a learned temporal module. A no-FaRL control collapses to 0.9%, showing that FaRL supplies semantic initialization while DINOv3 supplies dense spatial correspondence. The strongest correspondence appears at an intermediate layer: block 18 gives a 4.93x same-region versus cross-region discrimination ratio, compared with 1.48x at the final block. Against CLIP ViT-L/14, DINOv3 shows only a small aggregate advantage but a +16.8 pp gain on anatomical regions, indicating that image-level contrastive supervision captures coarse facial layout but not fine-grained anatomical identity. These results establish frozen DINOv3 as a strong zero-shot representation for region-level facial correspondence and identify intermediate self-supervised features as the most useful layer for dense face analysis.
The face is most accurately decomposed into features in the 18th layer out of 24—as if a prism produced the brightest rainbow not from the first ray, but deep inside the glass.
The sci-fi identification from 'Minority Report' becomes reality without colossal databases: a neural net, like an invisible spectroscope, dissects a face in an instant.