Generating High Resolution Images Using Transformers

Transformers are known for their long-range interactions with sequential data and are easily adaptable to different tasks, be it Natural Language Processing, Computer Vision or audio. Transformers are free to learn all complex relationships in the given input as they do not contain any inductive bias, unlike Convolution Neural Networks(CNN). This on the one hand increases expressivity but makes it computationally impractical for long sequences or high-quality images. In December 2020, Patrick Esser, Robin Rombach and Björn Ommer, AI researchers from Heidelberg University, Germany, published a paper on combining the Convolutional Neural Network(CNN) and Transformers to overcome the problem of producing high-resolution images: Taming Transformers for High-Resolution Image Synthesis. To m