DINO_V2_VITB14_768

DINO_V2_VITB14_768 is a medium image embedding model for strong visual grouping in EIDORA. Choose it when image similarity quality matters more than the fastest possible CPU runtime.

Best For

Higher-quality visual grouping and image retrieval.
Natural-image collections where semantic and visual structure both matter.
Projects that can trade some CPU speed for stronger embeddings.

Not Ideal For

Very large first-pass projects that need the smallest possible model.
Text, video, or audio inputs.
Identity, biometric, or other high-impact decisions.

Compute Tier

Medium: better quality or broader domain coverage with moderate runtime cost. Intended for recent laptops and desktops.

Inputs

image: required image input from media_source.

Output

The primary output is embedding, a float32 tensor shaped [batch, 768]. Embeddings are already normalized and are intended for cosine similarity.

Usage In EIDORA

EIDORA shows this package as a medium image embedding model in the Model Zoo. Use it for discovery maps, grouping, retrieval, and related embedding workflows.

Preprocessing

image: resize mode defaults to crop_center; rescale uses 1/255; normalize inside the ONNX graph with mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225].

Authorship And Citation

This ONNX package was produced by EIDORA from the original DINOv2 ViT-B/14 model. EIDORA converted the model to ONNX and is not the original model creator. Please cite DINOv2: Learning Robust Visual Features without Supervision and the original model repository when using this converted model.

Original model: https://github.com/facebookresearch/dinov2

Original paper: https://arxiv.org/abs/2304.07193

Authors: Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Aladin Bojanowski, Philippe Weinzaepfel, Hervé Jégou

@article{oquab2023dinov2,
  title={DINOv2: Learning Robust Visual Features without Supervision},
  author={Oquab, Maxime and Darcet, Timoth{\'e}e and Moutakanni, Th{\'e}o and Vo, Huy V. and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and Bojanowski, Piotr and Weinzaepfel, Philippe and J{\'e}gou, Herv{\'e}},
  journal={Transactions on Machine Learning Research},
  year={2024}
}

Training Data And Provenance

Base model: facebook/dinov2-base. Source repository: https://huggingface.co/facebook/dinov2-base. Known training data: LVD-142M, a curated dataset of 142 million images described by the DINOv2 authors. Package payload size: 346520584 bytes.

Evaluation And Validation

The package validation checks that the ONNX graph loads with ONNX Runtime CPU execution, runs the declared fixtures, returns finite float32 embeddings with the declared shape, and matches the artifact hash recorded in config.yaml.

Limitations And Safety

DINOv2 embeddings inherit the coverage gaps and biases of large-scale web imagery. They are not a reliable basis for identity, intent, protected-attribute, or other high-impact decisions.