In this issue:

• we dive deeper into Stable Diffusion; 

• we discuss retrieval augmented diffusion models that bring memory to text-to-image synthesis; 

• we explore Stable Diffusion interfaces. 

Enjoy the learning!  

💡 ML Concept of the Day: What is Stable Diffusion? 

Despite the progress in text-to-image synthesis models, there has been hesitation about open-sourcing many of those cutting-edge models. Most of the concerns are related to ethics and bias, given that images produced by those models can contain harmful or unethical content. Stability AI has been one of the AI startups that defied traditional convention and open-sourced its massively large Stable Diffusion model. As a result, Stable Diffusion has become one of the darlings of the data science community and the center of experimentation in the text-to-image generation space.  

Stable Diffusion is based on the latent diffusion architecture explored earlier in Edge#225. The core idea of this type of model is to replace the pixel space with representations that operate on a lower dimension but are equivalent in terms of information. Specifically, latent diffusion models operate in the latent space of large autoencoders, which reduces the computational complexity compared to other models. To achieve this, Stable Diffusion relies on three fundamental building blocks:  ...