ConvNeXt Explained: How CNNs Borrowed From Vision Transformers

Last Updated on October 29, 2026 by Editorial Team Author(s): Kasrasadatsharifi Originally published on Towards AI. Understanding the architectural decisions behind ConvNeXt, advantages and disadvantages over Vision Transformers and older CNNs. For years, CNN models dominated computer vision tasks such as classification, object detection, segmentaiton, etc. In 2020, with the introduction of Vision Transformers (ViTs) the focus shifted. They challenged CNNs dominance by introducing the global contextualization using self-attention. ViTs showed that self-attention can be highly effective in visual recognition but the question now is: was self-attention the only component contributing to their success or there are other more subtle factors like training strategies and architectural modifications also impacting their performance? After the introduction, the article explains what ConvNeXt changed and what it kept: it remains a CNN (with convolutional feature extraction, skip connections, and hierarchical stages) and does not use self-attention, but borrows ViT design principles in how it restructures convolutions and training. It describes key architectural modifications (replacing ResNet’s initial stem with a 4×4 stride-4 convolution, using depthwise 7×7 convolutions with pointwise channel mixing and an inverted bottleneck, and switching to layer norm plus GELU while separating downsampling from residual blocks). It argues that performance gains likely come from the combination of these changes plus modern training strategies applied to a ResNet-like baseline. The article then discusses limitations, especially around computational/memory efficiency, the difference between global self-attention and ConvNeXt’s more local/distant-region aggregation, and the relatively large model sizes for deployment. Finally, it presents practical experiments comparing ConvNeXt-Tiny, Swin-Tiny, and ResNet-50 on a smaller Oxford-IIIT Pets dataset and benchmarks inference latency/throughput, concluding that ConvNeXt can be highly competitive—especially for certain fine-tuning depths and single-image latency—while the best choice can depend strongly on fine-tuning strategy and batch size. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI