HLL-anon here. I want to clarify some things.
First, about tags. Could be useful for prompt list rentry and for
>>47322296 anon.
For hll models I mostly used images from gelbooru, not from danbooru. Note that some tags on these sites are different. I only did minimal tag editing.
for "copyright" tags: only shortest was kept (hololive english -> hololive). Which means tags like "hololive indonesia" do nothing.
for other tags - meta tags like "highres", "translated" - were removed
hll3.1 and below used booru tags only.
hll4 used booru tags + tags generated by WD tagger. Only longest tag was kept for repeating words ("shirt, black shirt" -> "black shirt"). It was a bad idea, i think.
Also hll4 had a lot of additional non-vtuber images, because I was trying to get better style and better backgrounds - see picrel. While it worked, it added some weirdness too. Issues with wrong wings are caused by those extra images.
Second, good luck to
>>47308568 anon with his merging experiments. Maybe he knows this, but I want to add that aesthetic scorers are pretty bad.
Basically, those scorers are trained that "text, flat colors, simple backgrounds = bad", "classical paintings, aquarelle colors, detailed backgrounds = good". They also hate asymmetry, so heterochromatic eyes will get scored poorly and they can't determine if character is "correct", only that colors and composition is good. You can't fully rely on automatic scorers.
Example of how images are scored -
http://captions.christoph-schuhmann.de/aesthetic_viz_laion_sac+logos+ava1-l14-linearMSE-en-2.37B.htmlThird. I'll probably post a new version of the model in a few days. This time - only big corpos, selected indies, and vtuber artists. Less additional images. Should be less weird and more accurate.
