EVA-CLIP: Improved Training Techniques for CLIP at Scale

Contrastive language-image pre-training, CLIP for short, has gained increasing attention for its potential in various scenarios. In this paper, we propose EVA-<PRE_TAG>CLIP</POST_TAG>, a series of models that significantly improve the efficiency and effectiveness of CLIP training. Our approach incorporates new techniques for representation learning, optimization, and augmentation, enabling EVA-<PRE_TAG>CLIP</POST_TAG> to achieve superior performance compared to previous CLIP models with the same number of parameters but significantly smaller training costs. Notably, our largest 5.0B-parameter EVA-02-<PRE_TAG>CLIP-E/14+</POST_TAG> with only 9 billion seen samples achieves 82.0 zero-shot top-1 accuracy on ImageNet-1K val. A smaller EVA-02-<PRE_TAG>CLIP-L/14+</POST_TAG> with only 430 million parameters and 6 billion seen samples achieves 80.4 zero-shot top-1 accuracy on ImageNet-1K val. To facilitate open access and open research, we release the complete suite of EVA-<PRE_TAG>CLIP</POST_TAG> to the community at https://github.com/baaivision/EVA/tree/master/EVA-<PRE_TAG>CLIP</POST_TAG>.