==========
2.0+
⬇⬇⬇⬇⬇
#Note, this is an SDXL lora, it will not work without an SDXL 1.0 checkpoint loaded.
Updated version using Juggernaut XL v7
I took a break from SD loras to focus on other ai stuff, (like local llms and learning python)
But I had to come back and try again when I saw how damn good Juggernaut has become for SDXL!
I've also at least doubled the dataset of images. (All actual images, no ai generated ones that I can tell) And manually curated most of the tags myself.
Honestly though, JuggernautXL does most of the heavy lifting and this lora only enhances "certain parts" ;)
Anyways, I hope some people get some use out of this and thanks for everyone posting what they have already been able to do with V1.
And here is the top 50 tags from the dataset you can use to bring out specific features in the lora:
1girl: 249
realistic: 240
solo: 240
nipples: 222
breasts: 220
nude: 192
navel: 172
long hair: 162
looking at viewer: 160
pussy: 146
medium breasts: 138
brown hair: 137
lips: 136
jewelry: 104
tits: 103
uncensored: 101
smile: 99
brown eyes: 81
completely nude: 77
small breasts: 70
black hair: 69
blonde hair: 65
spread legs: 64
barefoot: 63
standing: 62
shaved pussy: 58
vagina: 58
sitting: 56
mole: 54
piercing: 54
teeth: 54
female pubic hair: 50
indoors: 49
necklace: 49
blue eyes: 48
pubic hair: 47
grin: 46
navel piercing: 45
earrings: 44
large breasts: 44
outdoors: 42
short hair: 41
anus: 39
freckles: 38
feet: 37
lying: 37
nail polish: 37
topless: 35
ass: 33
breasts apart: 33
==========
1.0+
⬇⬇⬇⬇⬇
I couldn't find a lora to generate real enough looking natural female bodies and real skin.
So I made my own.
And as it turned out pretty good, I figured other people may be able to make use of it so I put it up here.
My goal was to get images as much like a actual photos as possible.
I used:
miamodel-SDXL_v30 as the base for my lora. (It's a great model)
64 images, and a lot of trail and error.
Notes on the model:
This lora is designed to appreciate the female body in all it's glory, and in it's natural form.
I'm sick of all the "insta filtered" photos where you can only see the absolute perfect image that they release after taking 100 pictures.
Tips:
I usually use it with miamodel at around 0.5-0.8 levels. Any higher and you start to get a lot of similar faces (and it can start to look cooked with models other than miamodel). But as with everything, feel free to experiment and have fun ;)
Adding moles or freckles tags should improve the "realness" of the skins texture. But you might want to turn them down as shown in the examples or they can just end up covered in them, and sometimes in weird places. lol :D
As you can see in the examples, I've lowered the weights on some of the tags, as at high weights the image will "zoom in" on specific areas. Experiment and see what works for the image style you're working on.
I know it's slow but after generating thousands of images I've found that with SDXL:
1024 x 1024 and 50+ sampling steps always works best.
If you're getting elongated bodies or weirdly shaped body parts using 768 x 1024, try stepping it up.
I'm also trying the new 3M samplers with A1111 and it seems like they may REQUIRE 50+ sampling steps sometimes.
V1.1 Update:
Highlights:
Higher detail all around (skin realism, body part detail etc)
Better variety of skin colors
Better variety in body types
Pubic hair keyword recognition (most of the time)
Better "spread legs" shots
Removed all text and watermarks!
Issues:
still has a bit of "same face syndrome" so make sure to add specific descriptors if you wanna see a bit more variety there.
"smile" "smirk" "grin" etc are still way over tuned so I turn them down to like 0.001
can still pop nipples through clothes sometimes
Thanks to everyone working on making more realistic models, ones that can actually understand what a human body really looks like. Ya'll are the real heroes <3
And special thanks to Aitrepreneur for his excellent and entertaining tutorial on how to make a lora: https://www.youtube.com/watch?v=N_zhQSx2Q3c
Without that, I might have never became motivated enough to actually make my own lora.
Made using:
sd-scripts (kohya's sd scripts)
by utilizing them through:
Description
Fir version, first (released) lora. May or may not make updates.
FAQ
Comments (12)
I'm still trying to transition from 1.5 to sdxl... but like, I'm noticing everything is significantly larger in size, especially LoRas. Is this really as small as one can make it with sdxl currently?
Sub 100MB is easily achievable, even though SDXL models will be bigger by default.
Realistically, yeah this is as small as I'm willing to make a lora right now.
if I make em any smaller, they start to lose details that I want to keep in the finished product.
But I'm still experimenting, so who knows?
Maybe someone in the future I or someone else out there can figure out a better way to keep details and have a small size too. Or maybe they've done it already and I just haven't seen it, lol :D
Either way, thanks for checking out my model, and have a great day!
I will say one thing though, if you're on the fence about XL.
Everything I've tried with XL has been better, easier to work with and higher quality.
But also slower and requires more GPU power and disk space.
I went and bought a 4090 just so I can work with all the AI stuff coming out lately.
It's worth it for me, but your situation and tolerance for generation speeds may vary 😀
@Light7799 That's because it's supporting higher resolutions. This is good and bad. Give it time and there will be better compression but there are limits. And as HackSlash said, XL has an initial learning curve (for everyone) but the results are much better. And once you get used to no longer having to make 800 word negative prompts and can using your normal narrative voice to create prompts, you'll see how much better it is. In a year there will be checkpoints that are so amazing at doing everything it'll blow all our minds.
But creators are having to work hard to keep up with demand; so I think we should donate and give where we can. Because it does take a lot of time to make these things for our enjoyment.
@HackSlash You mentioned power too - I limit my GPU to 300 watts. I did some testing and diablo 4 and games still push my card harder than stable diffusion does! SD renders are 250-290watts and gaming pushes past 310! Unlimited it hit 494 watts continuous (gaming) and 450watts continuous with stable diffusion. I never realized how much power gaming used and I'd let my game sit idle sometimes :O :O :O :O
@HackSlash If your curious about smaller model sizes, you can see the config I used here under "About this version". I know that it's going to depend heavily on the dataset though.
https://civitai.com/models/118841/hyper-bottm-heavy-sdxl
You can even use Kohya's resize lora script to reduce it by another ~50% with the sv-ratio option and a bit of tinkering. I did that for Hyperfusion and didn't have any meaningful concept loss.
Anyway good luck with training.
@throwawayjm Thanks for the tip. I didn't know about the resize lora script. I'll have to test it out when I have some time.
Thanks so much to everyone for checking out my LoRA!
Never expected it could get so many downloads so fast!
You all are awesome ✌
I'm working on a v1.1
Might be out in a couple of days :)
VERY useful Lora! :D
Thank you mate :D
I'm working on the next iteration right now.
It will be hopefully be better at everything the current one is good at...
But it takes like 6 hours to train per iteration, then I have to test it and see if it's actually an improvement.
So far I'm seeing better results in:
better leg spread shots
more diversity in faces
better variety of skin colors
And just all around better descriptions of the type of female body you want to be produced :D
Hopefully I should be finished with my tests by tomorrow (or at least enough to release a new update to this model)
And then I'm going to try and see if making the model a smaller file size. And see how that impacts the quality.
@HackSlash Oh, turned out I'm getting an error on this one, but I think it's on my end.
Are specific requirements for SDXL loras?
@joakimkunzdesign128 Without more info on the error, I don't think I can help much.
I use A1111 mostly, and there's plenty of guides out there to get it up and running with SDXL.
But here's a couple of things off the top of my head you can check:
Make sure you're running an SDXL checkpoint.
Make sure the lora is in the prompt and at the desired strength.
Use the "--no-half-vae" argument in A1111 when launching.
You could try re-downloading the lora file in case it got corrupted somehow.
But if all else fails, you can just re-install A1111 (or whatever you're using) from scratch and that'll fix just about any errors.