Deploying a text-to-image generation system

Design the job and artifact flow for a text-to-image service. Separate raw model weights from the complete pipeline, estimate image storage and transfer, and size generation capacity using measured workload throughput.

The reference is Stable Diffusion 3.5 Large, but the service also needs request handling, queues, artifact storage, and output checks. Begin with explicit workload assumptions, then trace how those components cooperate and fail.

Storage estimation

Assume 100 million daily active users, each requesting ten images per day. Separately assume 100 million stored profiles. Use decimal units and distinguish standing storage from daily growth.

• Generation-model weights: At two bytes per value, the assumed 8.1 billion parameters occupy 16.2 decimal GB. This is a raw parameter payload. The text encoders, image autoencoder, activations, and other runtime state increase the complete pipeline’s storage and memory requirements.

• Profiles: Assume 10 KB per stored account:

Scope: Keep model artifacts and profiles separate from daily image growth. Their sizes can change with versions, profile contents, and account count; they are not fixed for the lifetime of the service.