In this paper, we propose a custom Convolution layer, where kernels are predicted dynamically using a hybrid dense layer for each patch of the input image. This approach saves time and storage as the model doesn't have to remember kernels.