Taming Transformers for High-Resolution Image Synthesis

TLDR

Transformers excel at modeling long-range dependencies but lack local inductive bias, making them computationally infeasible for high‑resolution images. The study shows that integrating CNN inductive bias with transformer expressivity enables synthesis of high‑resolution images. The method learns a context‑rich vocabulary of image constituents with CNNs and then uses transformers to model their composition, enabling conditional synthesis guided by class labels and segmentations. The approach achieves the first semantically‑guided synthesis of megapixel images using transformers. Project page: https://git.io/JLlvY.

Abstract

Designed to learn long-range interactions on sequential data, transformers continue to show state-of-the-art results on a wide variety of tasks. In contrast to CNNs, they contain no inductive bias that prioritizes local interactions. This makes them expressive, but also computationally infeasible for long sequences, such as high-resolution images. We demonstrate how combining the effectiveness of the inductive bias of CNNs with the expressivity of transformers enables them to model and thereby synthesize high-resolution images. We show how to (i) use CNNs to learn a context-rich vocabulary of image constituents, and in turn (ii) utilize transformers to efficiently model their composition within high-resolution images. Our approach is readily applied to conditional synthesis tasks, where both non-spatial information, such as object classes, and spatial information, such as segmentations, can control the generated image. In particular, we present the first results on semantically-guided synthesis of megapixel images with transformers. Project page at https://git.io/JLlvY.