Generative diffusion models present unique operational difficulties due to their computational intensity. Moving these models from research environments to production systems that serve real users reliably and efficiently introduces significant engineering hurdles. This chapter focuses on understanding these specific difficulties.

We will analyze the compute and memory requirements inherent in the diffusion process, particularly during inference. We'll examine the trade-offs between generation latency (how fast an output is produced) and throughput (how many requests can be handled concurrently). Furthermore, we will review common system architectures used for serving large generative models, contrast synchronous and asynchronous processing methods for handling user requests, and adapt core MLOps principles for the continuous management of diffusion models in production environments. By the end of this chapter, you will have a clear picture of the main obstacles and foundational approaches for deploying diffusion models effectively.