Training is the expensive thing that happened once, somewhere else. Inference is the thing you pay for every time you press generate, and it is where a production budget actually goes.
It is worth separating in your head because the two have opposite economics. Training cost is sunk and enormous; inference cost is small, per-unit and entirely under your control through batch sizes, resolutions, step counts and how many failures you tolerate.
Every lever in this glossary that saves money saves it at inference.
What is the difference between training and inference?
Training builds the model, once, at enormous cost. Inference runs it, every time you generate, at a small per-use cost. Only the second one appears on your invoice.
What most affects inference cost?
Resolution, sequence length, step count and your acceptance rate. The last one is usually the largest and the least tracked.