I’m still figuring out what happens when settings are changed in stablediffusion. Tonight I played with sampling. With a really basic prompt, I increased the sampling by 1 going all the way to the max of 150, keeping everything else exactly the same. The results were interesting.
Note: I uploaded them in order, not sure why imgur changed that around. The number under the photo is the number of samples
Sampling is just how many passes Stable Diffusion makes to generate the image. It makes noise based on the seed and then diffuses it with the prior image in “latent space” as many times as you select.
The first few samples are going to be pretty abstract, but with each pass it can add more detail. At a certain point, additional passes aren’t likely to get much more detail, though, I think primarily due to the limitation of the resolution (512 for 1.5, for example). Instead, the new noise basically pushes the image in a slightly altered direction.
There’s no right answer to how many steps you should use. It’s going to vary between models, seeds, and your prompt. I find 20-40 works best in Stable Diffusion 1.5 based models.
If you generate an image that looks great other than some minor oddities, try adjusting the steps up or down a bit and see if it resolves it. If not, use a seed variation (the “Extra” checkbox next to the seed number in AUTO1111) to keep the original seed, but introduce slight alterations.
