Skip to content

Add ViT NEPA #42977

Description

@sbucaille

Model description

Summary : Next-Embedding Prediction: The Simple Secret to Strong Vision Learners

NEPA is a self-supervised method. It trains Vision Transformers to predict future patch embeddings. No complex loss functions or extra heads. Achieves 85.3% top-1 accuracy on ImageNet-1K with ViT-L.

The NEPA model got released two days ago and while the results are promising, I couldn't not notice how close to transformers the code they released was, and as @NielsRogge mentioned in his tweet, they heavily relied on the HF libraries.
So I suggest we could add this model to transformers and showcase how easy it becomes easy for researchers to pick and build on top of existing models. Anyway, I got a first draft PR to be opened.

Open source status

  • The model implementation is available
  • The model weights are available

Provide useful links for the implementation

Every useful links are in the HF paper page : https://huggingface.co/papers/2512.16922

@yonigozlan @SihanXU

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions