https://void.cat/d/6zJseJek9a6dAmmgsBZYUv.webp
It might helpful, or at least interesting, to know how a stable diffusion checkpoint was trained. I grabbed this script from reddit and packaged it up for easier use. It compares two checkpoints and shows the tokens with the largest difference between them. These are likely to be the words used more frequently during fine tuning.
The image was generated with 11 of the top tokens from the Photon checkpoint and a random prompt from the One Button Prompt extension. Three of the Photon tokens were: fancy steel mask, and the rest were nonsense words or word fragments: zzle elis eha fol hep spon abia wbo
The artists selected by One Button Prompt are important to the style, but the image style resembles neither of the artists.
I generated txt2img at 768x768, adjusted levels and curves in gimp, and brought it back to img2img for upscaling with ControlNet Tile and Ultimate SD upscale.
txt2img prompt:
zzle elis eha fol hep spon abia fancy steel mask wbo, art by Myoung Ho Lee, (art by Jacek Yerka:0.7) , landscape of a Maximalist Costa Rica, roots with Herb garden, at Midday, Ultrarealistic, Nostalgic lighting,
Steps: 30, Sampler: DPM++ SDE Karras, CFG scale: 3, Seed: 2631022029, Size: 768x768, Model hash: ec41bd2a82, Model: photon_v1, Lora hashes: "add_detail: 7c6bad76eb54, epiNoiseoffset_v2: d1131f7207d6", Version: v1.5.1
