As the name suggest, you get strip club babes with stripclub background and lighting.
Waifu or Blip style tags work, Example: Tittyclub, Woman posing
and then you can add, poses or race or face type of woman, or how many, you can do types of clothing and types of poses.
Example tags used in training: which were auto generated with blip captioning
Tittyclub, three women in fishnets sitting at a bar
Tittyclub, two women in fishnets posing for a picture
Tittyclub, two women in tights posing for a picture
negative prompting was done auto by Ruined Fooocus
Generation is not too restrictive you don't have to follow this formula 100%, once again in my trials, basic prompting or 1girl, posing on couch, braided ponytail, pink bikini
this type of tags also work.
Trained on civitai, on 500 images, I have another one trained on 1200 image with kohya ss on Google collab results are close. But I recommend and has been using the Civit trained more it does better with faces and clothing in my opinion, but Potentially the kohya ss have more poses or variation because it's trained on more images. I am posting both version so you can choose and pick.
I only use Fooocus MRE and Ruined Fooocus, I got the best resulting face with Ruined, but you will have to check yourself if the faces are bad or not, I don't know why MRE messed up the faces few times. SDXL and Sd15 are bad with fishnets, it tries, but the inaccuracies are clearer to see. Most of the models in the trained image had fishnet style clothing, It's not completely bad, but it does struggle with it.
Recommend, Jibmix or Juggernaut, Hassaku XL but that's because I tested a lot with those, but this is a concept lora so should work with most SDXL checkpoints, but realistic ones are better IMO, does not work with Pyros well, Pyros used to be my favorite probably needs an update or something.
I would like if the civitai trainer to allowed 1000 images with their corresponding captions at least, right now the upload is limited to 1000 files total, and it counts .txt, It's possible to upload 1000 images, but then you have to caption them by hand? Which I think is possible. But the trainer is really fast, I would prefer doing all Lora making on CivitAi rather than google collab, but the image and caption should be counted differently, so I can actually upload 1000 images and their caption files.
Description
Trained with Kohya_ss on google collab, trained on 1200 images, 10 epochs, about 7500 steps.
FAQ
Comments (5)
I didnt realize you could train NSFW content right her on Civitai, in XL too.. nice work man.
Am i crazy or does the civitai trained one look better
Thats what I think so too, well I used Kohya_SS on Google collab, using V100 which is 16GB VRAM, I am confident CivitAi trainer uses at least a A6000 which is 48GB VRAM, because CivitAi trainer is also way faster than V100, higher VRAM should mean better training.
Now I was thinking about using A100 to train on Google Collab, but that's going to run out the computation units very fast. I feel like I will just use CivitAI lora trainer from now on, for the 10 dollar I pay google, I can do same if not double the work and still be 3X, 4X faster. But I guess only downside is 500 buzz minimum, so better make 5000 step minimum loras lol
this lora amkes asses very big
This is a great lora, not just for stripclubs, but for adding a trashy stripcore aesthetic to nightlife, clubbing, and street scenes. It perfectly captures that raw, gritty vibe, making it versatile for various urban and nightlife-themed visuals.