No specific trigger word, but I recommend the following: "She has an innie pussy. Her pelvic region displays a closed labial structure." also if it should be in the shot "Her anus is visible." can help with that.
So this one took me quite a while. I had several goals in mind when creating this.
The first was no likeness contamination. 0 faces were used in the training data. The compromise there is you have to sometimes prompt in more detail to get the face in the shot since the training data was more close-up/cropped/pose focused, but it shouldn't be too difficult.
Next was flexibility in poses, angles, and skin tone. To achieve this I used over 200 images in the training data (and after lots of pruning and two previous training runs that had a few quirks I wanted to eliminate). Lots of testing for how krea 2 responds to prompts. Using that knowledge and captioning each one with a sophisticated LLM caption method I custom wrote mixed with feeding it specific concepts/poses with fixed phrasing and with manually reviewing and tweaking each prompt individually. Yeah it was a bit of a nightmare.
The next goal was photorealism. Not every single image in the dataset was high quality due to lack of diversity of high quality poses, but most are very high. I think I achieved the level of realism I was seeking though nothing is perfect.
Also, something that was important was every image must be a true "innie". No peeking of inner lips/clit whatsoever, including some very rare ones with her legs spread. I wanted to do this for consistency and I think I achieved that. The combo of realism and innie consistency is very difficult and I spent more time curating than captioning (and I spent over a week on captioning). Luckily I already had a very large library before I began the project.
I also included anuses and breasts in the anatomy and captioning them when they were present carefully.
What this model was not specifically trained to do: wetness, insertions, spreading (yes an innie looks slightly different in how the outer lip folds look when it's spread versus a non-innie). Insertions I know would be highly desirable and has its own look, but also naturally "opens" a bit. These features may fight with a typical closed innie though and would likely be difficult to caption/train. I may try to train a variation in the future with these features anyway, I may not. It's a ton of work and I'm doing it for free so yeah. Regardless, I wanted to get something strong and consistent out there, especially since krea 2 is so strong as a base for i2v models which may lack an ability to do good consistent innies on their own.
So how does this rank versus the other innie pussy lora's out there? Well, what this model excels at and what might set it apart from some others: pose variation, consistency, high quality realism, anus inclusion, skin tone variation. Check the examples if you want to see the poses and examples of how to prompt and where you may think it can potentially stand out from the competition.
If you spot any quirks, figure out any interesting use cases, or anything at else, I encourage you to comment them below. I may potentially even be able to address them if I make a new version in the future.
Description
FAQ
Comments (10)
Looks great! Thanks for all your efforts. ♥
I'm looking forward to the photorealistic version; this one makes the visuals look like something out of Disney.
I'd advise tweaking your prompt/workflow. Things like "1girl" and other tags, even something like "blue eyes" can pull krea towards cartoon very easily because it was trained on realism AND anime. Instead of blue eyes you should say "Dull blue-grey eyes" you have to apply that same principle to the whole prompt. I put my prompts into each example, that should give you a good starting base.
Here’s the thing: when I copy your prompt, I can reproduce the (photorealistic) style. But with my own prompt, when I don’t use LORA, I get photorealistic images; when I use LORA, they turn into a Disney-style look. I’ve also tried the Innie LORA model on the M99, but that didn’t change the art style either. Plus, I’m using natural language—I haven’t provided overly detailed descriptions of the characters, and I’ve tried to avoid anime-related tags as much as possible when describing expressions and scenes.
@suede2031691 Alright, considering you've seen that it works with someone else's prompt, I'll stress what I already said. It's your prompt, not my lora. Some prompts can sometimes produce toon and sometimes produce realism as soon as a tiny variable is changed. That is the temperamental nature of AI and a model with split training on both like krea 2. I think that's what's happening to you here. If you're expecting my lora to enforce realism no matter what, it won't do that. It's an anatomy lora, not a style one. That said, if you learn to prompt for what krea expects, you can make it behave more predictably regardless.
@suede2031691 Yes, your prompt does matter, but there's even an easier way to avoid the issue you decribe. Say goodbye to the Krea2 Turbo base model and use an NSFW checkpoint (e.g. SickOllie) or a photorealistic checkpoint that was also heavily trained on NSFW (eg. IntoRealism). I promise, you will never have that issue again! A second option, if you want to further use the Krea2 Turbo base model, use an additional NSFW lora such as Realism Engine, SNOFS or Mystic XXX. It does miracles!
Believe it or not, but you have just released the best innie pussy lora for Krea2 that ever existed!!
Thanks! Glad to see someone else who appreciates it.
@daldraeic It looks perfect, which means that it's not too big and the slit is not too short in frontal view standing poses. It does not produce unwanted pubic hair or pubic hair stubbles. It can be used in different poses and always looks good. It does not destroy or change faces which is very important for me since I mainly use character loras. It does not destroy skin details or makes the character to look like a doll. On the other hand It does not add unwanted or too much skin detail. It does not have conflicts with other loras used at the same time. And the filesize of the lora is not insanely big. The only thing that could still be improved is an option to also add the clitoral hood slightly poking out of the slit in cases if desired. This could be done with a second version of your lora. Although I need to confess that I haven't tried to prompt that yet. I don't know if your lora is able to produce a visible clitoral hood if prompted. I also haven't tested yet how your lora looks and behaves when the pussy is spread and if I would need to raise the strength for this purpose. Can you tell me something about that?
@terror_from_beyond122 If you can do any of those things it would be because of whatever OTHER loras or whatever the base model produces. There was 0 clits or spreading in the training data. They would conflict with the "pure" innie dataset if I included them, so if I made something like a "spread innie" lora it would intentionally be separate. At that point I'm really just having to find a bunch of dataset images that match a look which I don't have and would take likely weeks of time for how picky I am.