it seemed to have seen the (poor) audience and the background, but it seems like tenma blended straight into the picture, and wasn't able to make a consistent pose with her
which, kinda confirms my suspicion that my artstyle can't be read by AI at all, or at least, would struggle to read it
and of course, all my errors have her having exactly 3 of a limb
be it arms and legs in total, or 3 legs
i personally don't have a favorite number, but i do like primes because 13%, 41%(53% now iirc), 1/137, and whatnot tend to be very useful in approximating random probabilities and constants in the universe
first triplet: old pippa model
second triplet: new pippa model
third triplet: mixing both because i'm not getting what i want and at this point let's see how fucked it can get
<lora:Pippa:0.8> for new model, <lora:char-pippa:0.8> for old model, same values for both for the 3rd set
Negative prompt: (blurry, monochrome, worst quality:1.4, low quality:1.4, low detail, bad anatomy:2, bad res, error, malformed, cropped, pubic hair, censored, big lips, text, extra limbs, missing limbs), missing arms, missing legs, extra arms, extra legs, extra feet, extra fingers, missing fingers, anal, on side, futanari
i think pippa's just forever fucked to be like this, that body is fucked
gonna test a few more things, but i'll probably call it a day because it's a pain learning how ai works, because it takes forever to see the results if i'm experimenting with anything, from different tags, models, resolutions, samplers, steps, and so on
>kiki's model is so powerful, it turns the checkpoints that don't produce good results previously into good shit
or maybe the other models were just plain shit, what the fuck, why is this model so good
i feel like i've been playing with stone tools and finally got a CAD machine
it has its limitations in that it's the only lora that can be active, with others not being as effective, and i can't really figure out a way to get what i want with just prompting alone instead of relying a bit with a lora for positioning and details
but i feel like i've gotten closer to what i wanted
for some reason, this new model, although it can generate good stuff, is kinda bad
i'm not sure how to describe it, explain it, but i feel like i understand somewhat of how it works based on what it generates well and what it doesn't
the old model, while inconsistent and generally has a lot of errors and is prone to it, has more variety than what the new one seems to possess, or at the very least, it has a good enough abstract of pippa to have flexibility, while the new model is probably built to generate pippa holding guns if anything, given the page's preview images
the new model is very tight up the ass on a particular style, and maybe that's because my weights are too high, but i'll need to experiment more to test that
i spent nearly an entire fucking hour generating using DDIM as a sampler (instead of UniPC, which feels like it's more consistent in generating the good stuff) while creating a giant table comparing checkpoints and control modes
i think i gathered a good enough sample despite that; balanced and controlnet are what to go for if your lineart is peak, but prompt is more important to give the ai more freedom and let it do its thing
i'm trying to identify a way to hasten the video-generation process so i don't have to leave it open overnight for all the frames to generate, like reducing the steps needed to generate a satisfactory image, but i'm not sure how to hasten it
Add comment