This is a LoRA for puffy nipples. Who doesn't love those!? It favors smaller breasts, and it also has rather good "texturing" of breast skin and nipples. Just use the "puffies" keyword to help engage it.
I recommend a strength of approximately 0.7
Recently rebaked for SDXL! Give it a shot, post your results (it's my favorite part of making these is seeing other peoples ideas with them!)
Description
Initial version
FAQ
Comments (10)
love me some puffies.
I once helped someone move just because of their puffies.
I love that you did this, solid effort. Anything over 0.6 weight seems to get artifacts for me. I just this week started trying to train loras so I'm no expert but v.1 here seems a little overbaked? Keep at it, you've inspired me to have a go myself :)
Hey thanks Halilzeus!!! Much appreciated. Also thanks for the strength suggestion too, and the review. I'm also fairly new to this artform of training LoRA's. I'm happy to share my settings and whatnot, especially to learn more. I'm trying to create x/y grids to figure out if there's a like... earlier epoch or a specific strength. But I think I also need to play with steps-per-image and also batch size (which I used 100/image and a batch size of 6). I suspect I'll work on a v2 of this as well, and I have a couple other ideas brewing too!
A simple learning noob question, do this type of mod has to be in TI instead of a lora because they are very centric to a particular subject and can be directly used in prompt
I'm not sure if a TI is better than a lora for something like this -- mostly I just have been training LoRAs, so I used that technique for training this concept.
But this doc shows TI vs. LoRA pretty well (in two sections, but you get the idea) https://huggingface.co/docs/diffusers/training/text_inversion
You've inspired me... I might try training a TI using (a more limited version of) the same dataset, might be worth a try. I'm still definitely a learner, but I feel like this LoRA came out "ok" -- it's probably better to use for inpainting after you get the rest of your scene right, but, you might like the first generations from this LoRA too, depending on if you like everything about how I chose the dataset and how to train for it (which is still a WIP technique, as it is)
Thank for the reply. I asked for ti because I tend to understand the image composition like a hierarchy checkpoint -> lora -> TI. The checkpoint as backbone, the lora for the type of drawing and Ti for little details. And that depend also where the mod apply, in the backbone vectors or in attention vectors or in another technical area. It's complicated but seem essential to understand to create more precise composition.
Curious whether you used images where most of the woman (or at least face) was featured or mostly close crops of breasts? I had originally feared close crops would impact the photo too greatly, but have actually achieved pretty good results straight erasing the parts of the person I don't want to train on. But I get opinions all over the place so curious about your method.
Sorry so slow -- but it's a mix. Some with faces, some without. But I just selected a variety. I didn't specifically erase anything though. Hrmm, would be a cool experiment to do one without faces entirely, though. Good thinking, and thanks for asking.
To add my 2 cents, I have done training just of boobs and it works fine when you make faces. I believe the model relies on its main training data for things it already knows. Either way you should experiment, but I think the best method is to have a mix of faces/no faces so you get more detail in closeups. Just remember that EVERYTHING in the training data will be learned.