Bartosz Pampuch - Stable Diffusion under the hood
Topic: Stable Diffusion under the hood
Speaker: Bartosz Pampuch
The development of generative technologies opens up enormous possibilities for us.
The internet is full of guides and tutorials that introduce this world, presenting the perspective of the end user.
But how does it work underneath? Is it supported by a single network or a series of interconnected mechanisms?
How does Stable Diffusion learn images, patterns, and how does it generate them? What makes it capable of turning ordinary text into a picture? What tricks are used to make all this work efficiently?
During the session, I'll present the general architecture of Stable Diffusion and its key components. We'll talk about autoencoders, VAEs, probabilistic networks, forward and reverse diffusion, transformers, UNet architecture, cross-attention, ControlNet among others.