What Have We Learned from Pre-training Stable Diffusion Models on 2 Billion Images?

Stable Diffusion models enable exciting applications in media and creative industries. One can generate beautiful and unique images and use them in applications downstream. However, working with these models is very challenging.

Here, at Anyscale we worked closely with leading companies like Canva, Samsara, OpenAI and Jasper to enable their use cases at scale.

In this webinar, you will learn how to design and build an end-to-end system for training stable diffusion models (a text-to-image model family).

We will examine data processing and model training components and optimize their throughput and cost efficiency. You will gain in-depth knowledge on how to do this efficiently in order to handle massive images datasets while being mindful of the compute cost. You will gain practical insights on utilizing data streaming, heterogeneous clusters, and end-to-end system design.

Join this session to learn more about:

• Reference architecture for stable diffusion pre-training with Ray and Anyscale.

• Building an end-to-end solution for pre-training the stable diffusion v2 model on a massive dataset of approximately 2 billion images for less than $40,000.

• Eliminating preprocessing bottlenecks with Ray Data and improving training throughput by 30%.

Is this webinar right for me?

This technical webinar is especially useful for AI Engineers who want to explore ways to operationalize generative AI models at scale.

It is also useful for Infrastructure Engineers who plan to support GenAI use cases in their organizations.