Research map / Architectures and efficiency / CNN design
RepVGG: Making VGG-Style ConvNets Great Again
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the multi-branch training architecture with identity and 1x1 branches, then re-parameterize for inference (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 72.39 -> 75.14 | Helped |
| Add an identity branch (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 72.39 -> 74.79 | Helped |
| Add a 1x1 branch (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 72.39 -> 73.15 | Helped |
| Keep BN in the identity branch (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 74.18 -> 75.14 | Helped |
| Keep BN before branch addition instead of moving it after addition (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 73.52 -> 75.14 | Helped |
| Add ReLU to each branch before addition (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: +0.55 | Helped |
| Replace structural re-parameterization with DiracNet kernel re-parameterization (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 75.14 -> 73.97 | Hurt |
| Replace structural re-parameterization with trivial identity-kernel addition (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 75.14 -> 73.51 | Hurt |
| Replace RepVGG structural re-parameterization with ACB (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 75.14 -> 73.58 | Hurt |
| Replace the plain staged design with ResNet-like residual reorganization (ablation) | RepVGG-B0 · ImageNet-1K val | Accuracy: 75.14 -> 74.56 | Hurt |
| Use the RepVGG-A0 architecture instead of ResNet-18 | ImageNet-1K val, 120 epochs · ImageNet-1K val | Top-1 accuracy: 71.16 -> 72.41 | Helped |
| Use RepVGG-A2 instead of EfficientNet-B0 | ImageNet-1K val, 120 epochs · ImageNet-1K val | Top-1 accuracy: 75.11 -> 76.48 | Helped |
Closest papers by meaning
- Make RepVGG Greater Again: A Quantization-Aware Approach (AAAI 2024 · similarity 0.90)
- RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers (ICML 2025 · similarity 0.86)
- ReSprop: Reuse Sparsified Backpropagation (CVPR 2020 · similarity 0.85)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.83)
- GreedyViG: Dynamic Axial Graph Construction for Efficient Vision GNNs (CVPR 2024 · similarity 0.83)
- RepAn: Enhanced Annealing through Re-parameterization (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.