facebookresearch/dinov3

Features

• Self-Supervised Vision Representation Trainers - Implements large-scale self-supervised vision representation training using self-distillation and masked image modeling.
• Vision Transformers - Employs a vision transformer architecture that processes image patches as tokens using attention layers.
• Image Segmentation - Provides high-quality semantic image segmentation and foreground isolation using pretrained vision transformer heads.
• Visual-Textual Alignments - Maps visual embeddings to textual descriptions to support cross-modal retrieval and zero-shot vision tasks.
• Feature Extraction - Produces high-quality dense image representations and similarity maps for various vision tasks.
• Feature Extractors - Extracts high-resolution dense visual features and similarity maps from images without task-specific fine-tuning.
• Vision-Text Alignments - Implements an image-text alignment model that maps visual embeddings to textual descriptions for zero-shot recognition.
• Image Encoder Embedding Extractions - Generates vector representations of images using pretrained backbones via standard model loaders.
• Masked Image Modeling - Learns visual features by predicting missing image patches through masked image modeling.
• Self-Distillation Pipelines - Optimizes vision models using self-distillation to refine semantic and geometric feature representations.
• Teacher-Student Distillation - Utilizes self-distillation by training a student model to predict the output of a teacher model.
• Dense Embeddings - Generates high-resolution dense image embeddings and similarity maps to find correspondences.
• Zero-Shot Inference - Performs vision tasks without target class training by utilizing text-aligned model weights.
• Zero-Shot Segmentations - Isolates specific objects within an image without requiring training on those particular categories.
• Adapter Layers - Uses linear-head adapter layers to map frozen high-dimensional embeddings to specific labels for downstream tasks.
• Object Detection - Identifies and locates specific objects within images using pretrained detector heads.
• Monocular Depth Estimators - Predicts depth maps from single images by mapping pixels to distance values.
• Binary Segmentations - Isolates the primary subject of an image from its background to create binary masks.
• Distributed Vision Pre-training - Provides tools for distributed vision pre-training on GPU clusters to process massive unlabeled image sets.
• Image Classification - Categorizes images into predefined classes using pretrained classifier heads or linear evaluation methods.
• Vision Transformer Pre-training - Executes large-scale pre-training of vision transformers using self-supervised masked image modeling.
• Linear Classifiers - Categorizes images using extracted tokens and linear layers without extensive fine-tuning.
• Vision Model Training - Provides tools for distributed vision pretraining of self-supervised representations on GPU clusters.
• Semantic Segmentation - Maps image pixels to depth values or semantic labels using linear layers.
• Zero-Shot Segmentors - Isolates specific objects within images without requiring category-specific training.
• Cross-Modal Retrieval Training - Trains the model to align image representations with text embeddings for cross-modal retrieval.
• Vector Similarity Search - Identifies visually similar images by calculating nearest neighbors between representation tokens.

Star history