/vt/ - Virtual Youtubers

>>71029759
SDXL wheelchair is much more coherent
bonus: latest XL workflow with lightning and the latest version of diffusion-cg: https://files.catbox.moe/uafeg9.png

>>71030944
thanks

>>70986030
thanks for the lora names
they look great for dark

>>70986337
glad you liked

>>71005659
looks like it'll be worthwhile to pursue this further

I can think of a possible way to target something, but I'm yet to sit down and try code it
I'll share my idea, so if any anon here has the knowledge, they can quickly hack some code

Take stock CLIP-L from openai, use the vision encoder to encode an image of <<chuuba>>
Now take the CLIP-L from pony, get the text encoder output for all the tokens in the token dictionary, and do a similarity check
any random token with high similarity should be the hash

You can do the same with CLIP-bigG if you have the GPU for it
pre-cache the text encoder outputs as optimization

for art styles, use multiple images of the same artist and check common tokens excluding character attributes