Striptease thong with money bills in - an outfit concept. A winning bounty brought to you by @the_tdog Trained basically only on midbody shots so getting full body with face takes some "prompt engineering" (or rather, adjust token weights). Playing with LoRA weight seems to affect amount of money, so weights above 1 will give more bills but could break anatomy, while weights under 1 have more stable anatomy but money bills are scarce. Anyway, probably works best at 1.0. Example prompt would be "woman wearing money in thong".
Description
FAQ
Comments (24)
I see you got nice images with full face. Do you think they can be re used to train a better version of this LORA?
(by the way, I'm the one offering more buzz for the bounty, but no reply till now...)
IMO it's never a good idea to train a LoRA with AI output unless the original was already trained from the same source - it would generally only degrade whatever the original LoRA can do. So I would not recommend that, but it's technically possible indeed.
@dolirama126 Can you share the dataset you used? There must be an option in the menu on the right because I've seen someone doing this.
Edit: since there is a bounty active on this, better dont share the dataset. But if you have a tool to reduce size of images in batch, you can share the thumbs of the dataset.
@settima_ai Sorry, not sharing anything about the dataset while the bounty is active, but if you are interested even after it's over, remind me and I will let you know. As for the more Buzz, you can always tip the winner or the person whose entry you liked the most if the OP does not ever respond :)
@dolirama126 I didn't respond to the "more buzz" thing because I didn't want to get into complication about returning it if nobody entered. This is the first time I'd ever posted a bounty, and didn't really know how it worked. I'm still a bit clueless. But I think it's cool settima_ai offered more, and I'll be awarding you dolirama126.
@the_tdog Hey no problem at all, I'm fine with the original amount otherwise I would not post my entry in the first place. It was your bounty and you could do whatever you wanted with it - you really do not need to apologize for anything related to it, no harm was done. Thanks for the award, hope you will like the result :)
@dolirama126 Thanks for the entry, it's exactly what I was after!
I tested the LORA but now I understand what do you mean with "is not possible to reuse the AI images for the train"!
Infact, probably due to SD 1.5 limits, money came out always with certain grade of mess. Often is very messy.
The part I dont clearly understand is why money if often all around, and why is often used like it was a "underware" (money-bra and money-panty).
I suspect is SD thinking it's a cloth or something like that.
The success rate of the LORA is very low. I tryied some checkpoints, but I get 1 o 2 decent images over 100 generated: the girl is always perfect but the money is 95% of time missing, broken or in the wrong place.
Actually the LORA is barely usable for a "solo girl standing" like in your examples. Using it in a more complex scene would be a big pain for the impossibility to combine a pose with the money done right!
Thanks for the feedback. There is a reason why no one has done this concept yet and that is exactly because it's very hard to make it look "perfect", unless you use only images where money are all exactly at the same spot and never anywhere else. As soon as you try to use images where money are all around the thong string, AI will think it's possible to put them just everywhere and will do exactly that. If you need a better version, you can always try your luck and put up a new bounty - I have done what I was able to do with my skills and seeing how many people have already downloaded and liked this, I think my efforts were not in vain :)
@dolirama126 I had a flashback just now!!! Yesterday a guy told me he called "b4ll3tt" his ballet shoes as TRIGGER word because there often extra "balls" on the image, with SD 1.5 getting the part "ball" of the word "ballet". Once he changed the trigger word, problem solved!
In your case, the fact "money" is a word already known by most checkpoint may explain why some extra money if often seen around, flying or in the hand of the girl.
The fact the word "wearing" is already know too, may explain why sometime money is used to cover other parts of the body!
Please, consider the idea to use a different TRIGGER word, instead of "girl wearing money in thong"
something like: "girl w3ar1ng m0n3y in thong"
were "wearing" and "money" cant no more trigger undesired effects!
@settima_ai I am using no fixed trigger words, "money in thong" is just what I found to be working the best from the options I tested but you can be creative and try to experiment with different descriptions of the situation. Also you can try to play with word emphasis, like using (((money in thong))) or something similar - I think I have not tested that. And yes, it makes a difference because first I tested it with "thong with money" instead, but that was causing even more mess than the current phrase. I would say this LoRA can be pretty flexible and you might be able to find a phrase/description that would make this work even better than my recommendation :)
@dolirama126 Really? In this case, a new version using a separate Trigger word to activate the effect seems may be even better!
@settima_ai Let's hope someone else will come with a better version that will include a proper trigger then - I'm happy with this one and not planning to update it anytime soon :)
@dolirama126 I couldn't be happier. I'm getting pretty good results using Clip Skip 2, it's making the dollars look more like dollars.
@the_tdog Can you please explain how to set the "clip skip" to 2? I've heard about this many times, but no explaination about how to set and use.
I love seeing all the examples created by the LoRA I set a humble bounty for!
It sure seems to be pretty popular, you had a good idea for the bounty :)
Hey, I dunno if you saw my latest bounty request, but I'm betting if you made it, it'd be another popular one! Unhooking bra
@the_tdog Oh I haven't, not paying that much attention to bounties anymore but I might look at it. Anyway, good luck with that one too :)
@dolirama126 ya, you gotta follow where your interests lie... I just remembered you did a bang up job of it so I thought I'd give you a ping... I wish there was a "commission" section to Civitai, the bounties make it more of a contest than I'd like it to be. You don't take them, do you?
@the_tdog The whole bounty system will be reworked in the near future so there might be more features, including commisioning perhaps. I take commissions often, but my specialty are real people - concepts like these I'm only doing when I feel like experimenting, so I could not guarantee you anything unfortunately (it doesn't always work well as with the thong).
@dolirama126 would you be willing to give it a spin? I understand it might not work 100%, but I'll be willing to commision an attempt anyways
@the_tdog I can do it for free if you can send me a training dataset. I'm not sure how and where would I look for images for this concept, so that's the only real obstacle that would prevent me from trying to train this.
@dolirama126 no prob, i think I'd run into the same problem haha... how many images would be needed?