Stable Diffusion is an AI script, that as of when I’m writing this, can only be accessed by being in their Discord server, however, it should become open source soon. If you are in their Discord server, and want to make an image, but the settings are too confusing, this guide should help you make the best possible image with Stable Diffusion. Rather than just explaining how to use it, this guide also has lots of examples, so that you can see the effects of various settings.
Every single image will be generated using the prompt “A highly detailed 4K fantasy matte painting of city inside cave built around a long river. ArtStation, CGSociety, Unreal Engine, concept art, red and blue color scheme.” and all images will have the same seed:
Settings Comparison #1 Steps and CFG Scale:
Steps are how many times the program adds more to an image, and therefore is directly proportional to the time the image takes to generate. CFG scale is described as how much the script tries to match the prompt, but it doesn’t work well if too low or too high. In the chart below, rows represent the number of steps and columns represent CFG scale values.
Settings Comparison #2 Sampler and CFG Scale:
Before creating the table below, I do not know what sampler does, so I get to find out along with everyone else reading this guide. Rows are for CFG Scale, columns are for the sampler. Unlike the first comparison where both rows and columns relate, in this one, samplers are entirely different, rather than a range, so each column is a separate comparison, and one column does not relate to either column next to it any more than the furthest column from it.
Based on these two comparisons, it seems that steps over 50 don’t have too much effect, and all the samples are mostly the same, except plms with a high CFG scale.