A Study on CFG and Steps with Euler_a - a 494 image grid

[jlwinn8]

I've been very interested on the impact of step size and CFG on the Euler_a model. It's commonly noted that step size can completely change the image when using Euler_a. I started to notice a trend between CFG and Step Size, so I mapped it out on a giant grid. Here's what I learned:

• There's generally two areas on the grid that provided me with acceptable results - the green and the blue. The blue was more in line with what I was trying to produce, but that may just impacted by the prompt. So basically Euler_a can produce two "mini-seeds" of images. There was a random pocket of semi-related but unacceptable images highlighted in yellow, so maybe 3.

• There are two distinct "overbake and underbake" areas that produced completely unacceptable results. (highlighted in red)

• I had originally thought that a high CFG would universally "overbake" an image. However, overbaking can be completely remedied by increasing step size. With this experiment, a step size of 45 resulted in an acceptable image, even at a maxed CFG of 30.

• The purple images are very similar to the blue images, but I feel like they are slightly better and more detailed. That being said, step size needed to be relatively high, and I'm not sure it's worth the extra processing time.

More experimentation is needed. This is just a sample size of 1, but it generally followed my hypothesis.

[plaeground]

nice, thanks for sharing