Pyramid Sampler Testing

I'm using A1111 with advance_euler_sampler_extension. Currently, I use Euler_Max because it seems to work well with either simple prompts with low steps or complex prompts with higher steps. I've been pairing it with KL Optimal scheduler. I really have no idea what I'm doing, but I'm not going to let that stop me.

After updating, I saw that there were some new "pyramid" noise options. I quickly learned that these could not be used for the hires phase nor in-painting, but it sounds like this might be fixed in the future. For further testing, I'll use Euler_Max with KL Optimal scheduler for the hires and in-painting with Adetailer.

After some simple testing with my realism focused model, I settled on focusing on Euler_Pyramid (1, 2, 3). I ran another XYZ plot of all the schedule types for the initial generation. To get some variety in the plots, I use sd-dynamic-prompts with some prompts that I've developed a body of work with (1, 2, 3) and artistic styles as wildcards. This helps me understand fairly well what the results typically look like. KL Optimal, Align Your Steps seemed to have the best results which is consistent with some of my previous tests. Normal also seemed to do well.

I tested again with my illustration focused model comparing the three Euler Ancestral samplers in one XYZ plot. I think my prompt/wildcards led to some odd images, but the usual suspects ended up doing well (e.g., Uniform, KL Optimal, Beta, Normal).

I tried cleaning up the static part of the prompt to help the wildcard prompts and standardize the images by using fcDetailPortrait for only the first 10% (determined by scheduler) of the image generation using the transition syntax as well as alternating fantasy artist style prompts. This XYZ plot showed Align Your Steps was the only one that came out decently clean of major defects.

I think from these results I'll be shifting from using KL Optimal to Align Your Steps and I didn't see any major improvements in Euler Pyramid over Euler Max.

TL;DR: I'm sure I'll continue to tinker with other samplers, but for now Euler Max and Align Your Steps/KL Optimal will continue to be my goto sampler/scheduler combinations.

Future: I'd like to release in-painting versions of my models with my next release and maybe then I'll see some improvements with Adetailer fixing hands, but I haven't had much luck so far. It often creates more problems even when using A1111 Controlnet with Depth Hand Refiner. I think this is mostly because the Controlnet doesn't have enough context to the image because I can only set the mask dilation preprocessing to a maximum of 128, and I think it would likely need to see a lot more. Sure, I could in-paint them all manually, but my goal is to automate as much as possible as part of the original image generation.

Article image "diffused pyramid" was created using DALL-E because my GPU was busy drawing boobs for this article and CivitAI requires an image to save a draft. I'm also lazy and won't update it.