Update README.md
Browse filesMassive thanks to
@NicoGJ
and
@BetterThanNothing
for their amazing models Toonify and Prancing Invasion Ribbon Wave.
This Checkpoint (And Series of Checkpoints) started out of a distcontentment for the lack of really creative and stylized checkpoints on civit, theres some sure, but many of the truly awesome looking models lack certain capabilities, and many of the content packed models lack swagger. (in particular, i HATE that basic brown haircut, school uniform, suburbs background images that models spit out, ofc this is just the model generalizing but still) so i spent an unhealthy amount of time indoors during the summer and beyond quite literally planning out how im gonna make an actually fun to use. model, and on october 7th, combining the models mentioned above with overly specific merging, along with other loras and styles, i birthed this version of Bakas Mix... the dumb version
Trying not to inflate my own ego, i can say that this model turned out EXACTLY as i wanted, ill put a more specific pros and cons list below this so dont waste your time with my yap, but from long detailed prompts, to something thats simple and actually fun, this model can do it, and with a style thats fun and consistent. if your a casual you wont be stressed out with this one, and if your advanced you will love the intricate creativity thats offered. alright meat riding over, lets get on with the strengths, weaknesses, and settings
Strengths (more ego inflation)
-Prompting is fun asf. this model understands nuanced prompting like asking for something to be "unique, funny, attractive". character ideas can be as detailed as you want, but you can get stupid simple, asking for traits like "hippie, lazy, punk, horror, etc". you can use booru tags, and also relatively straight forward sentences
-Characters are in abundance. elsa, supergirl, venom, batgirl, asohka, bowsette, and so on. i dont think ive even scratched the surface of the characters the model knows (P.S. if you didnt get the desired character, try something like ((bowsette)). increase the strength for better results)
in addition to characters, character concepts like slime girls, witches, monster girls, villain, etc are light work
-Porn bro. i would go as far as saying this model can do any porn concept if prompted correctly. mating press, doggy style, pov, fingering, anal, you name it
-Pony Score Tags, i have never even used these in any image ive ever made with this model, maybe its better, maybe not, but they are certainly not needed
Weaknesses (trying not to be biased 😭)
-Long Distance people might be hit or miss, as you will learn from my settings below, i usually only upscale by 1.20x, so this might not even be a con, but rather fixed with a higher upscale, but people at a distance can look mushy, specifically eyes
-Two characters in detail together ofc. i've actually seen decent success with this, but of course this is not consistent
-Loras are hit or miss i think. i have not tested this much, could be wrong, but yeah. luckily in the somewhat futher future i have a fix for this and im super excited to work on this. stay tuned, but maybe dont because im lazy
-Holding Objects like Swords are a bit wonky, but certainly possible. i plan to also fix this, but with a few extra gens this wont be too harsh of an issue
Settings. All of the showcase images you see are completely stock with no post editing, no fancy comfyUI nodes, no controlnet, no inpainting. the only "post" thing i do is a latent upscale (hi res fix in forge and A1111). ive used ComfyUI (stock nodes), and Forge for the images seen, and copying the image into comfy/forge should give you the exact shit i did. i have also used the model in Fooocus, as well as sites like SeaArt, with no problems.
also, most images were first or second rolls, and i always gen 2 images at a time, generally all of the images shown were made in around 4-8 total images, when i needed to reroll, it was mostly due to the upscaling putting an extra finger, or removing a finger, as latent upscalea tend to do. sometimes i would also get a composition that didnt look exactly how i wanted (background wasnt as blurry as i wanted, character had clothes style i wasnt a fan of, stuff like that), and in some cases, a characters eyes would have artifacts, which is usually fixed by the upscale, but sometimes upscales take the flaw and details it lol.
for transparency sake i figured i would mention these. i can be confident in saying you wont need to reroll much due to awful anatomy, incoherent images, or artifacts
Steps: 30 or more, never tested under that, but im considering a lightning version of this. lemme know if u want that
Sampler: any of the DPM's with karras lol, im sure others will work fine
Negative Prompt: just used a copy paste i made years ago, as seen in the preview images, big negatives wont effect quality negatively, and small ones proly wont decrease quality much idk, just put watermark tags in there
Latent Upscale For ComfyUI: i use bislerp at the same Ksampler settings as your base image, and a denoise of 0.55-0.75, with a 1.20x upscale for speed
Hi Res Fix For Forge, A1111: I truthfully have not messed around much, but i use the standard 4xUltraSharp, with a denoise of 0.55-0.75, and an upscale of 1.20x. no other settings are touched
NOW AS FOR THE LISENCE, ITS A SUPER STRICT ONE THAT STATES.... just kidding, do what you want, the amazing creators of the models i based this model off of dont push a disgusting agenda
although, if you make something cool with the model, i would love to see it. not nessacery, but credit would be nice, mainly just so i have a way of finding the cool shit yall make.. and dont be weird or evil, your responsible for that
give the creators listed above buzz or a comment, like, money, mouth kisses, idk, something that helps
P.S. this will be on all of these places under my BakaBlitz name and model name: Civit.ai, Salt.ai, Tensor.art, SeaArt, HuggingFace, and any other place you ask for thats not super weird lmao