For another project I compared a set of models and wanted to share the results. High-res in attachments.
Clip skip 2, 30 steps, cfg scale of 7.5, k_dpmpp_2m sampler, and native square resolution. Exceptions for DreamShaper XL and Flux.
Prompt sources are from randomly chosen CivitAI images and are not further adjusted to each respective model to see how well they perform on "unsuited" input. Especially the anime/cartoon-oriented models obviously suffered from such prompts. I threw in SD1, SDXL, Cascade, and Flux models. Pony, for some reason, produced very poor results and is excluded. Prompt 4 on stable cascade consistently generated noise.