facebookresearch/dinov2

PyTorch code and models for the DINOv2 self-supervised learning method.

What it solves

DINOv2 provides robust, high-performance visual features that can be used for various computer vision tasks without requiring extensive fine-tuning. It eliminates the need for large-scale labeled datasets by learning visual representations from images without supervision.

How it works

It uses a Vision Transformer (ViT) architecture pretrained on a massive dataset of 142 million images without any labels or annotations. The project provides pretrained backbones (including versions with "registers" to improve feature quality) and specialized heads for tasks like image classification, depth estimation, and semantic segmentation. It also includes support for domain-specific backbones like XRay-DINO for medical imaging and Cell-DINO for microscopy.

Who it’s for

Computer vision researchers and developers who need high-quality, general-purpose visual features for downstream tasks such as classification, depth estimation, and semantic segmentation across different domains.

Highlights

  • Self-Supervised Learning: Pretrained on 142M images without labels.
  • Versatile Backbones: Offers multiple ViT sizes (Small, Base, Large, Giant) and versions with registers.
  • eluarkan a variety of tasks: Pretrained heads available for image classification, depth estimation (NYUd, KITTI), and semantic segmentation (ADE20K, VOC2012).
  • Domain Extensions: Includes specialized support for X-ray analysis and cell fluorescent microscopy.
  • Vision-Language Alignment: Includes dino.txt for zero-shot tasks via image- and pixel-level alignment.

Related

  • Project
    PaddlePaddle/PaddleFormers

    PaddleFormers is a PaddlePaddle‑based Transformers library that mirrors Hugging Face’s API while adding high‑performance distributed training (tensor, pipeline, expert parallelism, FP8, FlashMLA, etc.). It supports 100+ LLM/VLM families (DeepSeek, GLM‑4.5, Qwen, Llama, ERNIE‑4.5, Gemma‑3, etc.), full training pipelines (pre‑training, SFT/LoRA, DPO, function‑call, reasoning), and native support for Chinese domestic GPUs. Install via Docker or pip, then use familiar `AutoTokenizer`/`AutoModelForCausalLM` classes for inference or the `paddleformers-cli` for large‑scale training.

  • Project
    Farama-Foundation/Gymnasium

    Gymnasium is a community‑maintained Python library that provides a standard, versioned API for reinforcement‑learning environments (classic control, Box2D, Atari, MuJoCo, etc.). It’s a fork of OpenAI’s Gym and serves as the main platform for developing and benchmarking RL algorithms.

  • Project
    Lightning-AI/litgpt

    A high-performance framework for pretraining, finetuning, and deploying 20+ LLMs with from-scratch implementations and no abstraction layers.

  • Project
    DLR-RM/stable-baselines3

    A set of reliable PyTorch implementations of reinforcement learning algorithms designed to provide stable baselines for research and industry.