Text-to-image generative models

by Dr Luis Felipe Barba Flores (PSI)

Wednesday 12 Oct 2022, 11:30 → 13:00 Europe/Zurich

OFLG/402 or https://psich.zoom.us/j/68410446070?pwd=NkhFZ2dPQzVjT0NtNUxwNkl2MFhRQT09

Abstract:

The new king of image generation is here; in the last couple of years, diffusion models have displaced GAN’s  as de facto approach to produce realistic high resolution images. With the introduction of large models like Stable Diffusion, GLIDE, DALL-E2, and Imagen, denoising diffusion probabilistic models (DDPM) have shown remarkable superiority in conditional image generation. In this talk, we will explore the history of these developments, and provide some intuition about the mechanisms used to train them.