The ultimate goal of computer vision and learning are models that understand our (visual) world. Recently, learning such representations of our surroundings has been revolutionized by deep generative models that profoundly change the way we interact with, program, and solve problems with computers. However, most of the progress came from sizing up models - to the point where the necessary resources started to have profound detriments on future (academic) research, industry, and society.

This talk will contrast the most common generative models to date and highlight the very specific limitations they have despite their enormous potential. We will then investigate mitigation strategies such as Stable Diffusion and, more recently, flow matching approaches to significantly enhance efficiency and democratize AI. Besides visual synthesis, the talk will cover challenging applications such as the estimation of scene geometry.

Time permitting, we will also delve into the implications that generative AI is unfolding on our societies.