require comfy-ui custom node (>=0.0.4) : https://registry.comfy.org/publishers/ttddrr11/nodes/animadualkvstyle

should somewhat compatible with turbo and nag

test something to more easily overfit the style into anima model without text conditioning

similar to previous version, but attention now from the input latent instead of just forcing query to go through static kv

have to balance between --r_payload and --r_route, should use with --gradient_accumulation_steps or higher batch size so the model have more refined direction when train

code based on commit c162e90 of kohya-ss/sd-scripts repo on github

sometimes the style isn't showing, just bump the strength to 1.3 or something and it should overfit to show

how model were trained ( --r_route / --r_payload )

iel: ~300ish magazine + cosplay images, prompt mixed of natural languages and tags, 64/64 (barbie texture at best, not really good) (took me to this to realize using --gradient_accumulation_steps 4 helps the training a lot, all other was trained with effective batch size of 1 so it kinda noisy and may affect the prompt adherence)

itachi_3dt: mixed of manga colored panel and danbooru images, prompts are tags, ~10ish, 16/16 (overfit little too much)

soleilmtfbwy03: danbooru, , prompts are tags, ~10ish, 16/16 (16 was too small to correctly learn mapping the prompt, maybe should have bumped it to 32/* instead)

sakuram032: twitter images, , prompts are tags, ~20ish, 16/16 (artstyle change a little too much, losing capabilities too generate good hand, not like the base model was that good anyway, limbs proportion not reached the extreme of original artist)

reiji: cover images on danbooru, prompts are tags, ~10ish, 16/16

ceey: danbooru,prompts are tags, ~10ish, 16/16

daodtt: danbooru, prompts are tags, ~10ish, 8/16 (not to the same effect as the artist, but should be acceptable as a copycat)