dinov3-vitl16-pretrain-lvd1689m: Using Meta's Vision Foundation Model Without Fine-Tuning

A vision foundation model that extracts high-density feature vectors from image inputs. It uses the ViT-L architecture from the DINOv3 series developed by Meta AI and was pre-trained on 1.689 billion web images. It delivers competitive performance on various vision tasks, including classification, retrieval, and segmentation, without any separate fine-tuning.

Unlike conventional approaches, it is immediately usable with just frozen features. Simply attaching a linear layer or a k-NN classifier enables it to handle image classification, 3D keypoint correspondence, depth estimation, and video segmentation tracking. It is a model distilled using a ViT-7B model as the teacher, returning 201 tokens for 224x224 inputs.

Compared to DINOv2, it shows consistent performance in terms of geographic fairness and diversity, with particularly less performance degradation in low-income regions. The relative gap between Europe and Africa still exists but has improved over the previous version. Since fine-tuning reflects the bias of the training data in the features, using the base features as-is is recommended.

It can be loaded directly via the pipeline or AutoModel in the Hugging Face Transformers library. With a scale of 0.3B parameters, it supports inference in GPU environments. It is an open-source model that developers needing a vision backbone can apply immediately.