( ͡° ͜ʖ ͡°)
Note: The LORA name is "AbsurdlyGoodOW3DPornStyle", the example images use the development name.
V2:
Is a LORA again
Trained at 768x768, booru tag style
useful additional prompts: rim lighting, stylized 3D render, 3D, absurdres, sunny
Sometimes likes to generate multiple views/girls (might be tied to arhoangel prompt) or bad anatomy, put these in the negative prompt
Dataset consists of about 500 top quality Yeero and arhoangel renders and video frames (extracted one per animation via a script). Generally speaking solo renders were vertical, sex animations were horizontal.
still a bit finnicky sometimes but with these results i don't care, might further finetune it later
V2Beta:
Note: The V2Beta version is not as good as i initially thought. I'll try and get a proper V2 going but in the meantime i'll leave this version up for those who want to try it anyway.
it's a LoCon now, trained at 768x768
Much better at making 3D clothes and "coop gamemodes" now
suggested strength:0.8
best used with a character lora
rim lighting, stylized 3d render, 3d, overwatch, blender, arhoangel (or yeero) are heavily recommended prompts
Check the example images for pretty good prompts ( <lyco:LoCon2reps6000-000002:0.8> is indeed this LoCon, i renamed it a bit late)
Dataset is now about 1000 images
V1:
it's 2 AM here so i'll keep it short for now:
trigger word is overwatch 3d porn style
suggested strength: 1
blender, 3d are relevant extra prompts
trained on about 620 overwatch porn renders from a handful of top artists, lazily grabbed using grabber with: "[artist name] overwatch -animated"
some character LoRAs work best at around 0.8 strength, other at 0.5 (especially 2D anime girls)
rule of thumb: the more clothes your character wears, the less 3D it will look
defaults to an absolute fertility goddess body
also tends to turn blonde characters into Mercy
trained on nai, works pretty good on Anything V4.5, where my example images come from
check the example images for decent prompts
As usual with my stuff, lots of room for improvement but i like what it does already and should really just go to bed
Description
V2 LoCon needs more finetuning, but can already give some spectacular results. I recommend pairing this (0.8 strength) up with a character Lora (0.6 strength)
Expanded the dataset to include a single frame from from every animation in addition to previous regular renders. This seems to make it a lot better for "coop gamemodes" ( ͡° ͜ʖ ͡°)
V2Beta is a Lycoris/LoCon now, looks much better in certain situations, MUCH better at making 3D clothes
made with updated training tools, better trained in general, although i still have to tweak some parameters
FAQ
Comments (5)
The LoCon tended to botch things more than it really helped. I was running it on my ckpt merge which is pretty aight at 3D
Thanks for your feedback, i'll keep it in mind while tinkering further on v2.
uh hey, i really love your tf2 map LoRA so i decided to follow you and i got a notification for a pokemon style (gen 1-9) model and i can't seem to find it now, can you explain what happened to that?
Oh yeah, someone told me the exact same model already was on this site, so i deleted my reupload. I just couldn't find it because i didn't think to write it with an é lol. Sorry for that. Here's the model in question: https://civitai.com/models/26116/pokemon-style
@BelgOwned thanks dude, love your stuff.