• To create the newest art style of qiandaiyiyu, I've hand-picked 680 images down to 217 and made a curated dataset

• The dataset is up to date til 7.27.2024

• All drawings are near-original quality from 5 ~ 13MB in size

• Used wd-vit-large-tagger-v3 / wd-eva02-large-tagger-v3 taggers first to tag basic things, and then manually reviewed each one of the drawings to delete and add tags once more

• Background noises such as speech bubbles, text, artist name, patreon username, signature, twitter username etc are removed manually with photoshop

• Manually tagged camera angles such as from above, from side, from below and looking away, looking to the side, facing to the side etc.

• Due to the limitations of the variation of drawings, background images are severely limited. I'd suggest to use a Lora to enhance background or use another artist prompt.

• The artist's style sometimes bleeds belly elements and double eyelids. Recommend to use inpaints or photoshop if it's not going away in multiple image generations with negative prompts.

• Using upscaler is highly recommended

Recommended Strength is 0.7 - 1. Whatever you want, really.