The authors present RepVGG, a simple architecture whose inference-time body consists of nothing but a stack of 3x3 convolution and ReLU, while the training-time model has a multi-branch topology. The decoupling of training and inference architectures is realized by a structural re-parameterization technique, which converts the multi-branch training model into a plain inference model. RepVGG reaches over 80% top-1 accuracy on ImageNet with higher speed than ResNet-50 and ResNet-101, demonstrating that plain models can be made as powerful as complex multi-branch architectures.

Since the introduction of ResNet, modern convolutional neural network architectures have become increasingly complex with multi-branch designs such as residual connections, dense connections, and neural architecture search (NAS) derived structures. While these designs improve accuracy, they come at the cost of "increased memory consumption due to branch-level outputs" and "reduced inference speed due to the lack of support for efficient parallel computation." The classic VGG architecture with its plain stack of convolutions is "fast and memory-efficient" but has been deemed "outdated in terms of accuracy."

The authors argue that "the training-time and inference-time architecture need not be the same." A model can benefit from multi-branch topology during training (for better gradient flow and implicit regularization) while being converted to a plain architecture at inference time (for speed and simplicity). The key enabling technique is structural re-parameterization: "we equivalently convert a trained multi-branch model into a plain model using algebraic transformations on the convolution kernels."

Multi-branch architectures have dominated since ResNet introduced skip connections, followed by DenseNet's dense connectivity and NAS-derived models like EfficientNet. Model compression techniques such as pruning, quantization, and knowledge distillation aim to improve inference efficiency but "work within the same architectural paradigm." The concept of re-parameterization has appeared in kernel decomposition and Winograd-domain transformations, but prior work has not exploited it for converting between fundamentally different architectures.

During training, each RepVGG block consists of three parallel branches: a 3x3 convolution, a 1x1 convolution, and an identity shortcut (when input and output dimensions match). All branches are followed by batch normalization (BN), and their outputs are summed element-wise before ReLU activation. This multi-branch structure provides "the benefits of implicit regularization and better gradient flow similar to ResNet," resulting in higher training accuracy than a plain 3x3 stack.

The key insight is that convolution followed by BN can be fused into a single convolution with bias, and then multiple parallel convolutions of compatible kernel sizes can be merged into one. Specifically: (1) The 1x1 convolution kernel is zero-padded to 3x3. (2) The identity branch is represented as a 3x3 kernel with 1 at center and 0 elsewhere. (3) Each conv-BN sequence is fused by incorporating BN parameters into the convolution weights and bias. (4) The three resulting 3x3 kernels are summed to produce a single 3x3 convolution. This conversion is "mathematically exact — the plain model produces identical outputs to the multi-branch model for any input."

The overall RepVGG architecture follows a VGG-like layout with 5 stages, where each stage begins with a stride-2 layer for downsampling. The width is controlled by multipliers a and b applied to the base channel numbers. The authors propose several variants: RepVGG-A (with multipliers a=0.75-2.5, b=2.5) for lightweight models and RepVGG-B (with larger multipliers) for high-accuracy models. Notably, RepVGG uses "only 3x3 convolution as the only type of spatial convolution," which is "highly optimized on modern GPU and CPU hardware" through libraries like cuDNN and MKL.

On ImageNet classification, RepVGG-A0 achieves 72.41% top-1 accuracy at 1.36ms inference latency on a single 1080Ti GPU, while RepVGG-B3 reaches 80.52% at 5.01ms. For comparison, ResNet-50 achieves 76.15% at 3.07ms and ResNet-101 achieves 77.37% at 5.38ms. RepVGG-B1g2 achieves 77.79% at 3.46ms, outperforming ResNet-101 in both accuracy and speed. On semantic segmentation (Cityscapes) and object detection (COCO), RepVGG backbones show consistent improvements over ResNet counterparts. Compared to EfficientNet and RegNet which require depthwise or grouped convolutions, RepVGG achieves comparable accuracy with significantly higher actual throughput on standard GPU hardware.

RepVGG demonstrates that plain VGG-style architectures can compete with and even surpass complex multi-branch designs when equipped with structural re-parameterization. By decoupling training and inference architectures, the method achieves the "best of both worlds: the training benefits of multi-branch topology and the inference efficiency of plain models." The resulting models offer a favorable trade-off between accuracy and actual speed, and the simplicity of the inference model facilitates hardware optimization, quantization, and deployment.