Аннотация:In this paper we study the way to automate meme generation from textual prompts using autoencoders based on the Transformer architecture. For this we collected a dataset of about 5000 meme images. Then we run an OCR (optical character recognition) library Tesseract on top of these images to get English texts from them. We filtered poorly recognised texts using FastText language identification library to get only images with English texts and scores above 0.9. Then we trained image generation models using our dataset.