Roborobocap style - Lora版本v1.1 (ID: 44941)

Hi all. It's my first style Lora to emulate the style of a Korean NSFW artist Roborobocap (definitely check him out on pixiv, the art is lit). The lora was trained on all the images available online with Luriel model as a base.

BE WEARE, Lora has a number of huge problems:

1) I found it to be very strong: 0.5-0.7 is my range, if you go above it becomes risky.

2) I used images with text, so a negative prompt with (English text, text) is necessary.

3) Overall stability of outputs and quality of lora is kinda shit (hopefully v2-v3 will solve these problem)

I daily drive ComfyUI so you might not see prompts in civit interface (but If you also use ComfyUI just drop my pngs into the web interface and you'll get my workflow).

It is an nsfw lora and it is tailored for mature content, but you can get sfw images with it too.

Hope you all like it. If you have insights or your prompts are better then mine (which is likely) feel free to shit in the comments :)

描述:

Hey hey. I retreined Robo lora for Anime style models (AOM3 as base). It can also be used for realistic images, but just not as good. To me, this version looks more "pure" as a style model. I don't know what caused it. My educated guess is that AOM3 is just plain better for that type of art, so using it as a base model for training gives it an inherent advantage. I also removed most images with heavy use of text from the sample, so it is not necessary to include "text" in the negative prompt.

TLDR:

0) model now is NOT overfitted! You can use strength =1 as a default.

1) significantly improved diversity of poses, backgrounds etc.

2) model became more SFW (but it can generate good nsfw if you ask nicely)

3) style is detached from poses and characters - it should give you more control over the girl's appearance.

4) It works well with anime and realistic models, but realistic performance might be slightly less than in the first one, but it might be my skill issue when it comes to promoting.

训练词语: