Meta AI Releases EUPE: Compact Vision Encoders for Edge Computing

Meta AI has introduced the Efficient Universal Perception Encoder (EUPE), a new family of compact vision encoders designed to perform diverse computer vision tasks without the need for massive model architectures. By utilizing a unique distillation strategy, these models—all under 100 million parameters—successfully rival domain-specific experts in image understanding, dense prediction, and vision-language tasks, making them highly suitable for edge devices like smartphones and AR headsets.

The Challenge of Model Specialization

Modern computer vision relies on encoders that act as the "eyes" of an AI pipeline, converting raw pixels into feature vectors. Historically, these encoders have been highly specialized: models like CLIP and SigLIP 2 excel at vision-language tasks but struggle with dense prediction, while DINOv2 and DINOv3 provide excellent structural descriptors for segmentation but lack vision-language capabilities. For edge devices, deploying multiple specialized encoders to cover these different needs is compute-prohibitive, yet previous attempts to combine these capabilities into a single, efficient model through agglomerative distillation have often resulted in performance degradation.

A Three-Stage Distillation Strategy

The EUPE research team identified that efficient encoders often lack the representational capacity to absorb knowledge from multiple diverse teachers simultaneously. To solve this, they implemented a "scale up, then scale down" strategy. Instead of distilling directly from multiple experts, the team first trained a 1.9B parameter proxy model to unify knowledge from three distinct teachers: PEcore-G for classification, PElang-G for vision-language tasks, and DINOv3-H+ for dense prediction.
This proxy model then serves as a single, unified teacher for the efficient student models. The training process follows a three-stage pipeline: multi-teacher distillation into the proxy, fixed-resolution distillation into the student, and a final multi-resolution finetuning phase. This approach ensures the student models learn representations that generalize across spatial granularities while maintaining high efficiency.

Performance and Edge Deployment

The EUPE family includes models across both ViT and ConvNeXt architectures, all of which remain under 100 million parameters. In benchmarks, the EUPE-ViT-B model consistently outperformed domain-specific experts, achieving 84.1 on IN1k-KNN for image understanding and 52.4 mIoU on ADE20k for dense prediction. Qualitative analysis shows that EUPE-ViT-B successfully combines semantic coherence and fine-grained discrimination, qualities that were previously fragmented across different specialist models.
Designed for practical integration, the models are exported via ExecuTorch and demonstrate impressive latency, with the smallest variant, ViT-T, running in 6.8ms on an iPhone 15 Pro CPU. The research also highlights that data quality is paramount; models trained on the LVD-1689M dataset consistently outperformed those trained on the significantly larger MetaCLIP dataset, underscoring the importance of balanced, high-quality training data over sheer volume.