ARCHIVES

In this paper, we present DeepMeme, a system that uses convolutional neural networks (CNNs) and sophisticated natural language processing algorithms to automatically generate image-text memes. In contrast to conventional methods, DeepMeme incorporates a multi-modal attention mechanism that enables the text generation process to concentrate on an image's most prominent visual elements. To extract reliable picture embeddings, we specifically use a ResNet-based architecture that has been pre-trained on a sizable image dataset. After being refined on a carefully selected dataset of meme captions, these embeddings are subsequently input into a transformer-based language model, which allows for the creation of contextually appropriate and maybe amusing text. The method uses a new loss function that promotes stylistic consistency with well-known memes as well as semantic coherence between the image and text.