Oriel — AI & technology worth noticing, every day.

facebook/dinov3-vitl16-pretrain-lvd1689m

AI model · https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m

First seen by Oriel on .

AI summary

DINOv3 is a family of vision foundation models that produces dense features and outperforms earlier self- and weakly-supervised models across varied tasks without fine-tuning.

  • The family includes ViT and ConvNeXt backbones pretrained on web images or satellite imagery, with smaller models distilled from ViT-7B.
  • Models return class, patch, and register tokens; ViT inputs use 16-pixel patches and can support larger images whose dimensions are multiples of 16.
  • Frozen features support classification, retrieval, segmentation, depth estimation, object discovery, and video tasks, often with simple downstream classifiers.
  • Training combines DINO self-distillation, iBOT masked-image modeling, KoLeo regularization, and Gram anchoring; ViT-7B training used 61,440 Nvidia H100 hours and emitted 18t CO2eq.
  • The model card reports remaining geographic performance differences, including weaker results for low-income regions and a gap between Europe and Africa; fine-tuning may increase bias.

#computer vision #vision transformers #self-supervised learning #convnext

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 2,414, latest observation ; +1,016 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 693,430, latest observation ; -46,370 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Modelsfacebook/dinov3-vitl16-pretrain-lvd1689mhttps://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m Read discussion