PE-Core L/14

Overview

The Perception Encoder (PE) is a vision foundation model developed by Meta. It is designed as a general visual backbone for retrieval and related vision tasks. In this benchmark, it represents a generic open-weight image encoder rather than a domain-specialized product retrieval model.

Architecture

The L-14/336 variant used in this study is built on a Vision Transformer Large architecture with 336×336 image inputs. In this evaluation, it is used as a vision-only encoder that produces 1,024-dimensional image embeddings.

Capabilities

Perception Encoder is evaluated here as a general-purpose visual representation model for image retrieval, providing a comparison point against both specialized product models and managed multimodal embedding APIs.