transformers - 💡(How to fix) Fix [Model Request] Add EUPE (Efficient Universal Perception Encoder) by Meta AI [1 participants]

Model description

I would like to request adding support for EUPE (Efficient Universal Perception Encoder), a recent vision backbone released by Meta AI.

Paper: https://arxiv.org/abs/2603.22387 Code: https://github.com/facebookresearch/eupe

EUPE is a multi-purpose vision encoder trained via distillation from multiple domain-specific foundation models into a unified representation. It is designed to be efficient (edge-friendly), general-purpose (strong across diverse downstream tasks), and competitive with specialized encoders of similar size.

The model family includes both Vision Transformers (ViT-T/S/B, patch size 16) and ConvNeXt backbones (T/S/B) (see model weights on huggingface at facebook/eupe.

I would be willing to work on a PR for this model, including: model implementation (ViT + ConvNeXt variants), weight conversion script, tests and docs. Happy to align with maintainers on: scope (e.g., start with ViT only), naming conventions, integration approach.

Note that the non-commmercial research license for these models mirrors other Meta model releases (e.g. DINO, SAM) which already have implementations in the library.

Open source status

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

Provide useful links for the implementation

https://github.com/facebookresearch/eupe https://huggingface.co/facebook/EUPE-ViT-B @zcck3rn3l