Dinov2 Small Alternatives

dinov2-giant is the largest model in the DINOv2 family of self-supervised vision transformers developed by Meta. It produces rich image embeddings suitable for a wide range of downstream computer vision tasks. The model is available through the Hugging Face Transformers library and is widely used in research and production for image understanding.

Dinov2 Base is a base-sized vision transformer model trained using DINOv2 for image feature extraction. Hosted on Hugging Face by AI at Meta, it belongs to the class of vision transformers and is distributed with open weights under the Apache-2.0 license. The model supports direct use through the Transformers library. Code examples show loading it via a high-level pipeline for image-feature-extraction or by importing AutoImageProcessor and AutoModel from_pretrained with the identifier facebook/dinov2-base. It is provided in PyTorch and Safetensors formats. Intended for computer vision researchers and developers, the model is accompanied by a model card that includes a description, intended uses and limitations, how-to-use guidance, and citation information referencing arXiv:2304.07193. It can be deployed through Hugging Face inference providers, notebooks such as Google Colab and Kaggle, or local applications. The repository lists it among collections for image feature extraction and includes community contributions along with options to copy the model to a bucket or follow updates from the maintainer.

DINOv2 with registers is a base-sized Vision Transformer trained using the DINOv2 self-supervised method by Meta. It produces powerful image representations suitable for feature extraction, classification, and retrieval tasks. The model is available through the Hugging Face Transformers library.

This is the DINOv3 ViT-S/16 model pretrained on a large dataset (LVD-1689M). It learns powerful image features through self-supervised learning. The model can be used with the Transformers library for feature extraction and is suitable as a backbone for various computer vision applications.

The facebook/dinov3-vith16plus-pretrain-lvd1689m model is a self-supervised vision transformer available on Hugging Face. It belongs to the DINOv3 family and carries the specific ViT-H/16+ architecture. The model was pre-trained on the LVD-1689M dataset and addresses the need for general-purpose image representations that can be transferred to downstream computer vision tasks without task-specific labeled data. It performs image feature extraction. The model card lists it under the image-feature-extraction task tag and supplies example code that loads it through the Transformers pipeline for that purpose. Integration with the Hugging Face ecosystem occurs via the Transformers library, where users instantiate an AutoImageProcessor and an AutoModel. The repository provides both the model weights in Safetensors format and the associated code for direct loading or pipeline usage. Notebooks for Google Colab and Kaggle are referenced to support experimentation. The model originates from AI at Meta. Its paper is identified by the arXiv identifier 2508.10104. The license is explicitly named dinov3-license. The repository itself is hosted on Hugging Face, allowing deployment through standard inference providers or local applications that support the Transformers library.

This is a DINOv3 vision transformer (ViT-S/16+) model pre-trained by Meta AI on the large-scale LVD-1689M dataset using self-supervised learning. It excels at producing high-quality image embeddings and features suitable for downstream computer vision tasks such as classification, retrieval, and segmentation. The model is accessible via the Hugging Face Transformers library.

This is a Hugging Face repository hosting the DINOv3 ViT-L/16 pretrained model (arxiv:2508.10104). It provides a vision transformer trained using self-supervised learning for high-quality image feature extraction. The model can be used with the Transformers library for downstream computer vision tasks including classification, retrieval, and segmentation. It is intended for ML practitioners building on top of strong visual representations.