EVA-02: A visual representation for neon genesis

• Published in: Image and vision computing Vol. 149; p. 105171
• Main Authors: Fang, Yuxin, Sun, Quan, Wang, Xinggang, Huang, Tiejun, Wang, Xinlong, Cao, Yue
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.09.2024
• Subjects: Foundation model; Representation learning; Vision transformer; Representation learning; Vision transformer; Foundation model
• ISSN: 0262-8856
• DOI: 10.1016/j.imavis.2024.105171

We launch EVA-02, a next-generation Transformer-based visual representation pre-trained to reconstruct strong and robust language-aligned vision features via masked image modeling. With an updated plain Transformer architecture as well as extensive pre-training from an open & accessible giant CLIP vision encoder, EVA-02 demonstrates superior performance compared to prior state-of-the-art approaches across various representative vision tasks, while utilizing significantly fewer parameters and compute budgets. Notably, using exclusively publicly accessible training data, EVA-02 with only 304 M parameters achieves a phenomenal 90.0 fine-tuning top-1 accuracy on ImageNet-1 K val set. Additionally, our EVA-02-CLIP can reach up to 80.4 zero-shot top-1 on ImageNet-1 K, outperforming the previous largest & best open-sourced CLIP with only ∼1/6 parameters and ∼ 1/6 image-text training data. We offer four EVA-02 variants in various model sizes, ranging from 6 M to 304 M parameters, all with impressive performance. To facilitate open access and open research, we release the complete suite of EVA-02 to the community at https://github.com/baaivision/EVA/tree/master/EVA-02. •EVA-02, a plain Transformer-based visual representation, demonstrates superior performance in various vision tasks.•EVA-02 reduces model size through robust optimization, advanced activation functions, and position embedding.•EVA-02 achieves 90.0 fine-tuning top-1 accuracy on ImageNet-1K with only 304 M parameters.•EVA-02-CLIP outperforms the best open-sourced CLIP in zero-shot ImageNet-1K classification, using less training data.