Krea 2 is an incredible image model, and it's looking more and more like it will be the true successor to Stable Diffusion 1.5/SDXL in terms of mass adoption and post-release training (LoRA's, model finetunes, etc). But the base model and all the other LoRA's (so far) have only focused on the, uh.... roast beef sandwich aspect of the vagina. Lots of wrinkly meat folds (very pronounced inner vulvas). And if that's your thing, I got no beef with you.
But for me, I like a cleaner look. Smooth curvy lines that are proportionally plumped up to an aesthetically pleasing degree. Simple, basic, not complex.
---------------------------------------
VERSION 2 UPDATE: So far the #1 complaint about version 1 was that it created extremely smooth skin that lacked details (very plastic/3D looking). I have to admit, I had tunnel-vision when testing that LoRA and actually thought it looked fine. It's possible I did the same here, so you be sure to let me know so I can keep fine-tuning for future updates. I appreciate any feedback, even if it is negative, harsh, or just poking fun at how bad the LoRA is, lol.
Anywho, I completely overhauled the original dataset images by meticulously feeding every single one of them through an AI image editor (I used Klein 9B KV with a few good skin-texture/realism detailing LoRA's). I dialed-in each image as best I could (and to match the style of the original as best as possible) and trained at rank 32 (instead of rank 8). I trained to 3,000 steps like V1, but I actually found the 2,000 step version was much better (probably less overfitting). And that is the one I am providing here. I may train well past 3,000 to see if anything improves, and if it does, I'll do a 2.x version update.
The quality difference between V1 and V2 is night as day in terms of skin detail, but it's still not perfect. I may have gone a little too far on the skin texture detailing, as things can sometimes look a bit wrinkly or dried out (or like the character has some kind of skin condition). But hey, at least it's better than smooth non-detailed skin, right? Plus you can always adjust it down or up or mess with the samplers/schedulers to get exactly what you want. Euler/simple is the most balanced and euler/beta amps up the skin details. And if you do 12 steps on the sampler (instead of 8), it will add even more skin textures/details (usually too much, but you be the judge of that).
So while V2 is a huge improvement on realism/skin detail compared to V1, it is still not at the level of Krea 2 base quality. Maybe that is impossible, I don't know. But I did a lot of A/B testing with my V2 LoRA and the Krea 2 base, and the skin detail/realism was decently close (at least to my eyes). But my LoRA definitely does reduce the level of overall detail compared a baseline Krea 2 sample. Clothes might not be as detailed, there will be less stuff in the backgrounds etc.
Give it a try and let me know what you think. I've only done a very limited amount of testing (a few hundred images/prompts), so no doubt I am unaware of a lot of potential downsides (or maybe even some unknown upsides). Either way, for those who really hated that V1 was so smooth and lacked skin details, this should get a little bit closer to what you may want.
--------------------------------------------------
m99's Labiaplasty LoRA's are, IMO, the golden standard for innies, and I hope they eventually release something for Krea 2. But since training this model is surprisingly fast (and very effective), I thought I would give it a try myself. And I'm honestly shocked at how well this LoRA turned out (at least with the tests I've done so far).
Training Info:
My training dataset consisted entirely of synthetic images of women that I generated using m99's Labiaplasty LoRA with Stable Diffusion 1.5 (Cyberrealistic) and Flux.2 Klein 9B models. On all the images only the lower body (crotch/butt) were cropped/used, so there should be little (ideally none) influence on character faces/hair/body or even overall model styles or compositions.
This should stack with other LoRA's very well, though it will be diluted and will compete when stacking (especially if the other LoRA is also trained on vaginas), so adjust the strength of it and others accordingly to achieve a proper innie. From what I've tried with stacking, it almost always has a positive effect on the vagina even if it doesnt make it a true innie (plumping up the outer lips and reducing the overall inner vulva size). It's actually quite nice and allows for some great vaginal customization in this regard when combining with other vagina LoRA's or all-in-ones.
Usage:
I did not add a trigger word or any captions to this LoRA. To make it work, just enable and set the strength to whatever works best for you. It is trained at 1.00 strength. If stacking, you may need to adjust it higher (or adjust other LoRA's lower) to get the right results.
Description
Version 2 is a complete overhaul of the original dataset and I also trained on rank 32 instead of rank 8 (seems to help reduce the plastic look). I went through and spent many hours re-touching every single image in the dataset to add better lighting and most importantly, more realistic skin texture details. The biggest complaint of Version 1 was that it made skin very smooth and plastic/3D like, this version adds a lot of skin texture details to help nudge it closer to the base Krea 2 model details/look. And to be honest, it almost adds too much skin details if you use the beta scheduler (or dpmpp_sde/beta... this produces an almost horrific level of over-detailing).
FAQ
Comments (9)
v. 2 seems to be working great. Thanks for good work.
Thank you! This version is a bit on the experimental side. I did my best to maintain the original images from the V1 dataset. However the shape, curves and style of the innie did get altered a little bit for every single image in the V2 dataset (some more than others).
Luckily m99 does have an innie LoRA for Klein 9B, so that made a HUGE difference in maintaining the overall style as I made edits and added additional details.
But a lot still got changed (more than just skin details/textures). You can see this with the image I uploaded (the V1 vs V2 comparison).
It feels like a +3 realism, but a -2 innie. Some loss for more gains, I suppose.
Well done and it retains well against multiple styles too
I would say big progress compared to v1. However, skin texture is more of a synthetic patterned noise rather than anything realistic. If used at 0.7 strength it's bearable, anything more than that and we get a cooked image. I guess that's the plague of using a synthetic dataset.
For v3, I have two strategies to suggest.
The first one. For training the v3, use a two-part dataset. First, a mix of synthetic (50% smooth images and 50% of your V2 dataset). Second, a photographic dataset (high-quality close-up photos of large areas of human skin, e.g. thigh close-up, butt close-up and so on), tagged accordingly. I would start from a 50-50 mix of the 1st and 2nd datasets and see how it goes. If the second dataset is tagged accordingly (e.g. "abdomen close-up, skin texture, tanned skin, photo, raw photo", nothing else), it should learn the subject without overfitting.
Second suggestion: for your first dataset (the first 50% of smooth images), TAG THEM! Tag them "3d render, cgi, smooth skin" in addition to whatever they have now. The "skin" dataset would receive "photo, raw photo" tags. The processed synthetic dataset would receive neither tag (no "3d" and no "photo").
Another (better, more labor-intensive) strategy. You can do something else entirely - and that would be the better approach. Don't generate - INPAINT real photos! I mean, inpainting with Klein 9B (like shown here: https://civitai.red/images/130286425). What I mean is: get a few close-up photographs (high-quality photos) and edit them in Klein 9B with that m99 LoRA and "fix her crotch" prompt. This way you'll get a high-quality dataset that has both the skin (real skin! not synthetic!) and the subject you want. I think you won't even need many; just 10 to 30 images is really all you need to teach that concept. You can complement this dataset with a few "3d, cgi, smooth skin" samples, too.
Your last suggestion is actually a really good idea, I hadn't even thought of doing the inpainting approach using real images with Klein and m99's LoRA. That way the only synthetic skin would be the vagina (and m99 in Klein has pretty good skin details). Thanks for the idea, I'm definitely going to try that one for version 3 and see how it turns out👍
Update after more testing: severe "sameface" issue detected. With v2 at 0.7 strength, all generations tend to converge into a very similar "default" face. Did your dataset include any faces at all?
Nope, no faces or even upper body. Just cropped crotch/backside butt areas. So something else must be causing this for you unless this is some other side-effect of the training.
MY SAVIOR!!!!!! THX THX THX!!!!!!
"tunnel-vision"
Me too...meeeee too ;)
Great Lora thank you!