A ConvNet for the 2020s introduces ConvNeXt, a family of pure ConvNet models that compete favorably with vision Transformers (ViTs) in terms of accuracy, scalability, and efficiency. The paper explores the design space of ConvNets and tests the limits of what a pure ConvNet can achieve. By gradually modernizing a standard ResNet toward the design of a vision Transformer, the authors identify key components that contribute to performance differences. ConvNeXt, constructed entirely from standard ConvNet modules, achieves 87.8% ImageNet top-1 accuracy and outperforms Swin Transformers on COCO detection and ADE20K segmentation while maintaining the simplicity and efficiency of standard ConvNets. The paper also evaluates ConvNeXt on various vision tasks, including image classification, object detection, and semantic segmentation, demonstrating its competitive performance. The results show that ConvNeXt can achieve similar or better performance than vision Transformers without the need for specialized modules such as shifted window attention or relative position biases. The paper also discusses the inductive biases of ConvNets and their advantages over vision Transformers, highlighting the importance of convolution in computer vision. The findings suggest that ConvNets can be as effective as vision Transformers in many tasks, and that the design of ConvNeXt provides a practical and efficient alternative to vision Transformers.A ConvNet for the 2020s introduces ConvNeXt, a family of pure ConvNet models that compete favorably with vision Transformers (ViTs) in terms of accuracy, scalability, and efficiency. The paper explores the design space of ConvNets and tests the limits of what a pure ConvNet can achieve. By gradually modernizing a standard ResNet toward the design of a vision Transformer, the authors identify key components that contribute to performance differences. ConvNeXt, constructed entirely from standard ConvNet modules, achieves 87.8% ImageNet top-1 accuracy and outperforms Swin Transformers on COCO detection and ADE20K segmentation while maintaining the simplicity and efficiency of standard ConvNets. The paper also evaluates ConvNeXt on various vision tasks, including image classification, object detection, and semantic segmentation, demonstrating its competitive performance. The results show that ConvNeXt can achieve similar or better performance than vision Transformers without the need for specialized modules such as shifted window attention or relative position biases. The paper also discusses the inductive biases of ConvNets and their advantages over vision Transformers, highlighting the importance of convolution in computer vision. The findings suggest that ConvNets can be as effective as vision Transformers in many tasks, and that the design of ConvNeXt provides a practical and efficient alternative to vision Transformers.