For a long time I treated generative models like a vending machine: insert prompt, receive artifact. The results were fine and forgettable. The shift came when I started treating the model the way Monet treated a pond — something to return to, again and again, under different light.
A diffusion model denoises toward an image the way a painter resolves a canvas: coarse masses first, then the smaller decisions that carry all the feeling. If you intervene early you change the composition; if you intervene late you only change the surface. Knowing which is which is most of the craft.
So now I build interfaces that let me intervene at the right altitude — nudging the structure while it is still molten, then stepping back to let the texture settle. It feels less like prompting and more like glazing.