Grippy Pussy
Details
Download Files
Model description
Trained on the grippiest of grippy pussies, after tedious work I can say, without ego, that I've hit it out of the park with this one. This is my first LoRA and I already consider it my magnum opus, that's how well it works. This LoRA is solely based around a from behind doggystyle/cowgirl POV. It can create a healthy variety of grippy pussy types, and also works well across SDXL models with a lot of modularity. The cover image isn't even the best example of how grippy things can get with this LoRA. It adds quite a sense of realism to images.
Of course, you need an insertion of some type for a pussy to grip, so this LoRA requires that to be specified for best results. It can create insertions on its own, but it's best to specify those details, along with other details like body type, hair color, pose, image quality, lighting, etc.
I will not be gatekeeping this project behind a buzz paywall, mostly because I want people to actually enjoy getting to use it, despite all the effort and time it took to make. I hope people appreciate that, and if you can, throw as much buzz my way that you can part with through the tip option near the top of the page.
_________
Recommended strength: 0.8 - 1.0
Trigger word: GrippyPussy
_________
Training insights: This model is technically undertrained, deliberately. The reasoning behind this is because you want the overall structure of the concept, without overly adhering to the other details of training images. So it's not actually undertrained in a sense. This means the LoRA is very modular in the ability to prompt for differences in insertion type/size, body type, etc. It also means the LoRA plays well with the usage of other LoRA. It benefits from detailed prompts and LoRA that affect things like style/lighting/body/skin/insertion type and size.
I've trained additional versions, and they can somewhat perform better at low strengths, but it loses a lot of modularity and variety, bleeding too much of the original training data into them, and at that point it could be considered to be overtrained. After all, we're not trying to train on asses or the details of the insertion object, we're trying to train on a very specific part of anatomy that is generally a small part of the image, while also keeping it flexible enough to have variety.
In any case, this LoRA as is, can produce great results, but if you want to up the game on it a bit, things like adding a few more steps or mixing with other LoRA and using High Res Fix or Kohya Deep Shrink can benefit the final result. I wouldn't really go over a strength of 1, you could maybe push it to 1.2, but generally the typical LoRA strengths of 0.8 to 1 work great, with 0.8 giving good flexibility and variety. It can be run at lower strength like 0.6 with some decent results for a more subtle effect, and that might just be down to personal preference.
Also, DMD2 LoRA is pretty much a must for SDXL at this point. It's soo superior than running checkpoints without it, that there's no reason to not be using DMD2 for everything.
Captioning insights & prompting tips: This was captioned with JoyCaption, so treating the word like a character name might help with prompting. For example "GrippyPussy's
To make it further clear, let's define what encompasses "GrippyPussy's" domain: The whole ass area, pussy, and insertion. That's largely what makes this LoRA soo versatile. It's not trained on any other attributes, therefore those aspects can be more modular.
Of course you don't really need to prompt this way, but I think it does open up some clever ways to prompt. It can potentially add to the possibilities in prompting more specific details. It might seem easier to just go in the captions and change things around, but base SDXL isn't quite NSFW on its own, so it will pick up on the word "vulva" more than "pussy" when it comes to training on the base model, and that's important if you want a concept to be recognized when training with base SDXL, as that will strengthen its results both in base SDXL and community created NSFW SDXL checkpoints.
At the same time, text encoding does somewhat latch onto combined words as both words individually. For example, if you prompted "MeowingCat", you'd probably generally get a cat meowing, or at least a picture of a cat, just like if you prompted "GrippyPussy" without using the LoRA, you'd probably still generally get a NSFW image of a pussy, so strategic use of trigger word design can sometimes benefit concepts by controlling how unintended language aspects bleed into it, which can be somewhat better than just using a trigger word of numbers replacing letters like you tend to see many models use. This allows the LoRA to better utilize and mesh with similar existing concepts in a checkpoint.