Keynote

“Compression in the Age of GenAI: Perfect Realism at Every Bitrate”
Dr. Lucas Theis

Abstract:
New insights in neural compression, AI, and related fields suggest an impending paradigm shift in lossy data compression. Generative methods are now capable of generating content that is nearly indistinguishable from uncompressed data, even at extremely low bit- rates. Transcending rate-distortion theory, these approaches force us to rethink fundamental aspects of lossy compression such as the role of distortion. They also challenge traditional methods of evaluation which have focused on distortions and low-level statistics. Some new coding schemes even deviate from the ubiquitous transform coding setting in fundamental ways, offering a new perspective as well as improved performance. Other advances in lossy compression with much smaller and faster neural-network-based decoders suggest that we will also finally be able to bring some of these developments into practical codecs. In this talk, I will give a high-level overview of these exciting new developments in lossy compression. I will delve into image compression using diffusion models as one example, and highlight important open questions that remain.

Slides: PDF