Dyna-DINO: Efficient ViT Distillation Via Adaptive Representation Anchoring

Paper | GitHub

This repository contains the ViT-Tiny and ViT-S checkpoints (No Register) distilled from ViT-G DINOv2 on ImageNet-100 and ImageNet-1K. The knowledge distillation process follows the procedure proposed in the paper "Dyna-DINO: Efficient ViT Distillation Via Adaptive Representation Anchoring".

Introduction

Vision Foundation Models (VFMs) with ViT backbones, such as DINOv2, are computationally demanding. LEAP (Layer-skipping Efficiency via Adaptive Progression) is a training curriculum for ViT feature-based knowledge distillation. Instead of supervising the student against a fixed teacher block, LEAP advances the supervisory target through the teacher's feature maps (shallow-to-deep) based on online CKA alignment. This allows the student to build a foundational representation before tackling higher-level abstractions.

Use cases

The ViT models output feature maps that can be used for a variety of downstream tasks, including:

• Image Classification
• Instance Retrieval
• Semantic Segmentation

Performance

ImageNet-100:

ImageNet-1K:

Citation

@article{DynaDINO,
  title={Dyna-DINO: Efficient ViT Distillation Via Adaptive Representation Anchoring},
  author={Zhang, Jiaqi and Lee, Ashton and Wong, Anthony and Zou, John and BuGhanem, Sami and Balestriero, Randall},
  journal={arXiv preprint arXiv:2606.19483},
  year={2026}
}