This LoRA creates beautiful lingeries. You can control how revealing it gets and which parts of lingerie to generate. Version 1.0 has edited 0.2's database, manually cropped images, redone tags and was trained with better settings. It's database contains over 130 images (768x768, without faces). If faces generates poorly use ADetailer.
Trigger prompt: sexy lingerie
Lingerie set prompt: sexy lingerie, garter belt, thighhighs, thong, bra
Pasties prompt (don't use "sexy lingerie"): "pasties, panties" or "pasties, heart pasties, panties" (you might need few tries)
Other useful prompts: from behind, bodysuit, fishnet, see-through, collar
To change color use addictional prompt: [...], [your color] lingerie, sexy lingerie, [...]
Recommended LoRA weight: 0.7-1.0
You can vote or give your idea for next trained model here! I'm gonna try to train most popular ones.
If you liked my lora consider buying me a Coffee or giving this model 5* and posting your creations😊
v0.2 description
A lora that upgrades woman's lingerie. You can control how revealing it gets and which parts of lingerie to generate. For now I’ve created database containing over 140 images.
Use the LoRA weight between 0.8 and 1 for the best results.
Prompt: sexy lingerie
Other useful prompts: panties, bra, lingerie, thong, thighhighs, garter belt, high heels, see-through, swimsuit, bikini, pasties, lace, cupless bra, bow, choker, fishnets, topless, gloves, pantyhose, no panties, bodysuit.
I recommend using ADetailer extension otherwise faces on you images might generate poorly! You can install it directly from Extensions tab in A1111 WebUI. Default settings works fine.
v0.1 description
A simple lora that upgrades woman's lingerie. You can also control how revealing it gets.
Use the LoRA weight between 0.8 and 1.4 for the best results. Higher weight makes lingeries more revealing.
Prompt: sexy lingerie
For example: full body shot, a photo of a beautiful woman standing in lingerie, blonde hair, garden, sexy lingerie, <lora:sexylingeriev0.1:1>
Description
- version 1.0 - database created with over 130 images (768x768, without faces), probably final version
FAQ
Comments (11)
could I learn about the training set of this model??
Well, it's just a lot of good quality photos of lingeries
I recommend looking up my training settings by clicking "i" button on the lora in A1111 WebUI. I've also had results like this on higher epoches (at least sometimes), so you need to be careful not to overtrain the model (still while using 2 or more of my models at the same time often same problem appears).
@olaz thank you very much Master, I'm new and I didn't know about the info 'i' on the model, I will look and do other tests. on the last epochs by lowering the intensity I have whole person results, but I still have a long way to go, thanks again for your time
@Romanowvv You're welcome and good luck :)
@Romanowvv I also use wd-v1-4-swinv2-tagger-v2, then I add trigger tag and I edit other tags a bit with dataset tag editor Web UI extension. I don't use "head out of frame" tag though. I think the problem isn't in the tags, it is in the training. Try using slow training rate, it should let it learn enough about the clothes and still won't chop off heads :)
There is an easy way to use loras with faces trained in them without them affecting the face (or affecting just a tiny bit). I plan to write a quick tutorial for that. It's basically using ADetailer extension (or inpainting face manually) with changed positive prompt (without lora in it) and ~0.4-0.6 inpainting strength. This gives really good results and you don't have to bother cropping images.
@olaz Thank you so much MASTER, I'll put everything to the test. Invaluable advice. as you said I had already adopted ADetailer or roope for my face, the influence of training can sometimes also be felt on skin colour, body type and hair. but I'm improving the datasetting and the tag phase and it's already better, now I'm at 0.12% of the result I'd like to obtain. the phase of eliminating some tags from the images in the dataset is not yet clear to me, especially with regards to clothes. if my goal is to train a bikini: should I remove everything related to the garment (bikini) from the photo tags? Here I find many opinions on this and all conflicting. Now I'm taking a middle path. but I still have to do a lot of testing. Any of your tutorials would be greatly appreciated, not just by me. thank you for your time.
This is a great Lora! Would it be possible to share the config you used to build it?
All training parameters are saved in lora's metadata. To access it, in A1111 WebUI click the "i" icon next to the lora model.