i had already jumped into stable diffusion with both feet at this point. i probably would not have gotten this deep without automatic1111's stable-diffusion-webui. there were and absolutely still are ways to generate images via command line but i very much do not prefer command line over a half-decent graphical user interface. i was having way too much fun trying different prompts that other people were sharing mid-september and started finding out that there were models other than the base sd-v1-4.ckpt i had grabbed a week before from CompVis' Hugging Face repo.
different models produce different outputs based on the same input. the difference might be small or it might be insanely huge. so i needed a way to test the difference between models as a way of kinda knowing if i might like like it or not. i have a background in photography, i never have and never will consider myself much of a photographer but i worked in photo labs for many years. not the kind of photo labs you see at wal-mart or drug stores but the kind that could do just about anything you wanted on site without having to send it somewhere else (generally called sending to an outlab) and wait days or weeks for whatever you order to be returned. in fact, the photo lab i worked in was the outlab for many wal-marts and drug stores for several states. there were even some companies (factories, offices, whatever) that we had contracts with where we'd put a physical dropbox somewhere at their location for thier employees to drop their film off and get photos and slides returned within a day or two.
anyway, getting a bit off the point there, but because of my background i was already familiar with the need for and a method or two of standardized testing that would always produce the same, very similar, or at least predictable results when some of your variables changed like film stock, paper manufacturer, chemistry, etc. Shirley cards or negatives was a coloquial term used by my boss back then, and since this is a thing i'm pretty familiar with i went with the same idea. my reference prompt for testing different models would be "pretty lady". I generated a couple hundred "pretty lady" images at 512x512 with random seeds using the stable diffusion 1.4 model and decided on one that i thought looked pretty good compared to all the others. The seed for my shirley test is 470304.
this would prove useful for testing different sampling methods as well. Here's two that i still have saved where i was comparing (using X/Y/Z plots) SD1.4 against SD1.5 and a waifu diffusion model mix:
