Next-Embedding Prediction Makes Strong Vision Learners

TL;DR

This paper introduces NEPA, a simple yet effective self-supervised vision learning method that trains models to predict future patch embeddings, achieving strong ImageNet and segmentation results without complex auxiliary tasks.

Contribution

Proposes Next-Embedding Predictive Autoregression (NEPA), a novel embedding prediction approach for self-supervised vision learning that simplifies architecture and training.

Findings

Achieves 83.8% top-1 accuracy on ImageNet-1K with ViT-B after fine-tuning.

Effective transfer to semantic segmentation on ADE20K.

No need for pixel reconstruction or contrastive loss.

Abstract

Inspired by the success of generative pretraining in natural language, we ask whether the same principles can yield strong self-supervised visual learners. Instead of training models to output features for downstream use, we train them to generate embeddings to perform predictive tasks directly. This work explores such a shift from learning representations to learning models. Specifically, models learn to predict future patch embeddings conditioned on past ones, using causal masking and stop gradient, which we refer to as Next-Embedding Predictive Autoregression (NEPA). We demonstrate that a simple Transformer pretrained on ImageNet-1k with next embedding prediction as its sole learning objective is effective - no pixel reconstruction, discrete tokens, contrastive loss, or task-specific heads. This formulation retains architectural simplicity and scalability, without requiring additional design complexity. NEPA achieves strong results across tasks, attaining 83.8% and 85.3% top-1 accuracy on ImageNet-1K with ViT-B and ViT-L backbones after fine-tuning, and transferring effectively to semantic segmentation on ADE20K. We believe generative pretraining from embeddings provides a simple, scalable, and potentially modality-agnostic alternative to visual self-supervised learning.