Intro

This is an HTML page that helps to see how a ckpt model KNOWS about specific TAG

If you don't see any clue about a certain TAG, that might indicate that this ckpt model doesn't handle the TAG well.

Tags

You can see tags from :

• (250412 updated) Face & emotions

  • https://danbooru.donmai.us/wiki_pages/tag_group%3Aface_tags

• (250412 updated) Sex positions v2 ( with extra sex test tag )

  • https://danbooru.donmai.us/wiki_pages/tag_group%3Asexual_positions

• Camera related

  • https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition

• Sex positions (v1)

  • https://danbooru.donmai.us/wiki_pages/tag_group%3Asexual_positions

• Posture

  • https://danbooru.donmai.us/wiki_pages/tag_group%3Aposture

• Gestures

  • https://danbooru.donmai.us/wiki_pages/tag_group%3Agestures

Params

• X

  • seed : 1, 2, 3

• Y

  • positive prompt : best quality, TAG

  • negative prompt : worst quality, bad quality, worst detail

Base ckpt model used

• wai v13 -- WAI Illustrious v13

Tips

• Download & unzip the zip file, find index.html and open

  • You can click on image to preview

  • Size of cell could be toggle via the panel at the bottom right

• The script use for generating such benchmark grid would also be published later

  • I don't like the XY tool in webui for benchmarking a LoRa that's trained. ( it requires a lot of manual work )

    • So I wrote a tool to auto-run a benchmark based on pre-set config

  • Any other tool you find useful, please recommend~! 🌷

Findings

• 1) I think some TAG would be triggered better with other related tag/s, instead of being 'TAG' alone by itself

  • CLUE : spooning (sex position) should work, but none of the image give me a correct image (only 1 girl in the image). I think 1girl, 1boy, sex, spooning will work better?

• 2) The base model (wai i13) tends to draws 1girl by default when quality tags are given