Kolmogorov–Arnold Networks (KAN) have recently gained significant attention. We test whether KAN can be used for computer vision tasks. In this paper, we evaluate Convolutional KAN (CKAN) on image classification and object detection. Our experiments reveal that CKAN demonstrates promising performance in certain architectures, such as Visual Geometry Group 16-layer network (VGG16) and Residual 18-layer network (ResNet18) on Fashion Modified National Institute of Standards and Technology dataset (FashionMNIST), where it achieves higher accuracy with fewer parameters. However, in other configurations, such as ResNet18 and You Only Look Once 11 s-version (Yolo11s) on Canadian Institute For Advanced Research-100 dataset (CIFAR-100), while accuracy improves, the network complexity also increases significantly. For smaller networks like LeNet and MobileNet, the integration of CKAN does not consistently deliver stable improvements; in some cases, it even leads to a slight decrease in accuracy. In our object detection experiments using Yolo11s as the baseline, we explored two approaches: direct replacement of convolutional kernels with KAN-based kernels and integration with depthwise separable convolutions. However, neither approach consistently outperforms the original Yolo11s model, and performance can sometimes degrade. These findings underscore that simply replacing or adding KAN structures does not universally guarantee performance improvements. A careful balance between accuracy enhancement and network complexity is essential when incorporating CKAN into various architectures.