DINOv3
A self-supervised visual pre-training model based on the Vision Transformer architecture, belonging to the DINO series. It is pre-trained on large-scale image data through self-distillation to learn general visual feature representations, and can be adapted to various downstream vision tasks.
Timeline (last 90 days)
The SPOOK experiments evaluated OpenCLIP and DINOv3 models on the ImageNet and Flickr30k datasets.
On DINOv3, a blank canvas after sparse pixel edits likewise showed high target alignment and near-top prototype rankings.
The off-manifold failure mode exists in both language-aligned multimodal encoders (OpenCLIP) and self-supervised vision encoders (DINOv3), suggesting that this vulnerability may be a shared property of the representation geometry of modern vision encoders.
DINOv3 is a representative self-supervised vision encoder that learns visual representations on unlabeled images through self-distillation and does not involve text supervision.
All Facts (4)
The SPOOK experiments evaluated OpenCLIP and DINOv3 models on the ImageNet and Flickr30k datasets.
50%UnverifiedOn DINOv3, a blank canvas after sparse pixel edits likewise showed high target alignment and near-top prototype rankings.
50%UnverifiedThe off-manifold failure mode exists in both language-aligned multimodal encoders (OpenCLIP) and self-supervised vision encoders (DINOv3), suggesting that this vulnerability may be a shared property of the representation geometry of modern vision encoders.
50%UnverifiedDINOv3 is a representative self-supervised vision encoder that learns visual representations on unlabeled images through self-distillation and does not involve text supervision.
50%