No artist name used as a trigger word

• No trigger word required

• Helps maintain the latest style

• Reduces issues involving image quality

• Prevents the model from calling on older art styles

• The previous LoRAs should had never been trained on artist's name

Dataset

• Watermark manually edited out

• Primarily new images with a smaller portion of older ones included

• Used to be higher but ended up with 312 imgs, did go over them a bunch. Yes, still a lot but, these are pretty consistent in style

• lucy/rebecca influence brought down to increase variety and avoid same-face syndrome, at least to a certain degree

• On some images missing this, manually added the subtle “line” shading (around nips) that are frequently displayed in some of their art works. This was done to increase likelihood of appearing, just because... it's hot 😎😭

Training details

• Model trained on Illustrious XL 1.0 and WAI v15

• Illustrious XL 1.0 compared against training on Noob eps and Illustrious v0.1:

  • Illustrious v0.1 produced the weakest results

  • Noob performed well but had occasional issues with color saturation and image quality

  • Illustrious XL 1.0 delivered great results while avoiding the problems seen in Noob

  • No more pony v6, it's old and f them score tags fr. v7 maybe, previews do not give much confidence