RepVGG is a simple yet powerful convolutional neural network architecture inspired by VGG, featuring a VGG-like inference-time structure composed solely of 3×3 convolutions and ReLU layers, while the training-time model uses a multi-branch topology. This architecture is achieved through structural re-parameterization, allowing the model to switch between the two topologies. RepVGG achieves over 80% top-1 accuracy on ImageNet, outperforming many complex models. It runs significantly faster than ResNet-50 and ResNet-101 with higher accuracy, showing a favorable accuracy-speed trade-off compared to state-of-the-art models like EfficientNet and RegNet. The model is efficient, easy to implement, and suitable for deployment on GPUs and specialized hardware. RepVGG's plain topology makes it memory-efficient and fast on general computing devices, and it can be further optimized for specialized hardware. The model's architecture is defined by depth and width, with different variants such as RepVGG-A and RepVGG-B. Experiments show that RepVGG outperforms ResNet and other models in image classification and semantic segmentation tasks. The model's structural re-parameterization technique is crucial for training effective plain ConvNets, and it is verified through ablation studies and comparisons with other models. RepVGG is a simple and efficient architecture that achieves high performance with a favorable trade-off between accuracy and speed.RepVGG is a simple yet powerful convolutional neural network architecture inspired by VGG, featuring a VGG-like inference-time structure composed solely of 3×3 convolutions and ReLU layers, while the training-time model uses a multi-branch topology. This architecture is achieved through structural re-parameterization, allowing the model to switch between the two topologies. RepVGG achieves over 80% top-1 accuracy on ImageNet, outperforming many complex models. It runs significantly faster than ResNet-50 and ResNet-101 with higher accuracy, showing a favorable accuracy-speed trade-off compared to state-of-the-art models like EfficientNet and RegNet. The model is efficient, easy to implement, and suitable for deployment on GPUs and specialized hardware. RepVGG's plain topology makes it memory-efficient and fast on general computing devices, and it can be further optimized for specialized hardware. The model's architecture is defined by depth and width, with different variants such as RepVGG-A and RepVGG-B. Experiments show that RepVGG outperforms ResNet and other models in image classification and semantic segmentation tasks. The model's structural re-parameterization technique is crucial for training effective plain ConvNets, and it is verified through ablation studies and comparisons with other models. RepVGG is a simple and efficient architecture that achieves high performance with a favorable trade-off between accuracy and speed.