Prompt engineering is the iterative design of prompts and settings to steer Stable Diffusion toward better images. Generate, inspect, tweak, and repeat; “engineering” is less blueprint and more educated prompt hacking. section 1.3 section 1.0
A practical workflow:
- Prefer clear, descriptive terms over poetic wording when the goal is a specific visual result. section 1.3.1
- Change one or more prompt terms, negative prompts, or generation settings, then compare the resulting images. section 1.3 section 2.0
- Record the seed when testing changes so results remain reproducible. section 1.0 section 1.3.1
- Study prompts recommended for a checkpoint because different models respond differently to keywords and phrases. section 5.2.2
From poetic coffee to a clearer prompt
“A damn fine cup of coffee, black as midnight on a moonless night” can produce inconsistent results. Replacing it with “A cup of black coffee” gives the model more direct guidance; generate several images, inspect them, and keep refining. section 1.3.1
Figure 1.18 A straight forward prompt yields more cups of black coffee.

Compare the straightforward coffee prompt with the earlier poetic attempt to see how wording changes the visible results. section 1.3.1


















































