The YARN (Yet Another Realistic NSFW) Collection includes diverse styles for use with Krea 2. While these Loras can be used one at a time, we recommend mixing different Loras (strength ~0.2 - 1.5) to combine their strengths.
Unless otherwise specified, pictures in the linked post are generated with strength 1.0, and only one Lora, which differs greatly from recommended use.
Here is a quick description of the style intended with each lora, starting with the latest releases.
Hegre
Nothing is superfluous in this austere look on erotism.
ShowyBeauty
Who ordered the most beautiful women on the internet?
BlackedRaw
Same vibe as Blacked, but at night with a blinding flashlight in the performer's face.
GFRevenge
It's not about getting even, it's about remembering the good times that are past.
Throated
These women have the pornstar look, and do not care about their mascara getting smudged during messy blowjobs. Pictures in the linked post generated at strength 1.5.
Phil Flash
Take a trip back to the 2000s with these playful cuties. Short skirts, ultra tight low waisted denim, tiny plain thongs are all on the menu.
Tiny4K
These tight vertically challenged women take on girthy cocks for your viewing pleasure.
Ersties
These German lesbians are fully deconstructed and ready to take you into their world.
BrattySis
Thansksgiving, Christmas, Easter ... Every holiday is an excuse for unforgettable family memories. But coming back home to your step sister can wake some tensions. "Help, step bro! I'm stuck in the dryer."
AbbyWinters
Discover the best of the bush with these Australian amateur lesbians.
InTheCrack
Heavy makeup and shiny bubble butts for those who don't fear looking down the barrel.
Vixen
Glamour and provocative, these models pose around the villa or at the beach with "that" look, before giving themselves to their audience.
Blacked
Bold and beautiful women face shiny BCCs in front of studio lights with a polished professional aesthetics. Starts with high end lingerie or sporty underwear with the signature belt, end often ends up with a cum shower.
FTV
Beautiful young woman lift up their dresses in public, undress in front of the camera for the first time, and will make you blush the next time you're in a room with a cucumber. Side effects include upskirts, pussy spreading, and deep insertion of sextoys.
Tushy
The sun shines on the glistening bodies of these beautiful woman, but not where they like to take their cocks. In this primarily anal focused experience, the ultra-rich bang indoors and outdoors, near the pool or in the sand.
ThisYearsModel
Fall in love with these young mischievous women. Simple, warm and beautiful.
SexArt
The prettiest women know they are the prettiest. Mostly Eastern European woman wanting to make you feel good.
MetArt
These nude artistic pictures have soft edges for a dreamy look on the most flawless woman.
Zishy
Candid photographs of beautiful amateur girls next doors.
Description
Heavy makeup and shiny bubble buts for those who don't fear looking down the barrel.
FAQ
Comments (6)
Great work, how big was your dataset? Do you have the budget to do Minimax H3? 😀
Depends on the version. Always on the larger side for each iteration (10k-80k). Some iterations even have only one epoch. I'm not really interested in video yet, sorry.
@TimmyHodor That's ok, you've done fantastic work. One epoch with tens of thousands sounds great for generalisation. Would you be able to explain rough budget, software, settings, etc? You must have learned a lot along the way!
@ShinyLilly Sure!
Hardware: Everything local. 4090 + 64GB RAM
Software: Joycaption beta 1 for captioning. Onetrainer for training (the best by far IMO). 768pxl x 768 pxl is enough. batch 4 fits well on my 4090. AdamW, beta1=beta2 = 0.995 to 0.998. # of iterations : 6k to 25-30k. lr 1e-3 with alpha=1.
Dataset preparation: Either only get ultra high definition base images (before onetrainer auto resize) ok use Klein9B img2Img to remove jpeg compression artifacts (takes a while). Else, the final result is almost always blurry.
For style, I do recommend a high number of images. For character, way fewer is ok (100 I guess).
Not much more to say I guess, unless you have a specific question.
@TimmyHodor I also appreciate you sharing your setup (soft and hard). Regarding character traing I was suprices how good Krea loras got with fewer pics, depending on the pics of course..my smalles one is only 11 pictures and amazingly good, normally I am between 30-60 pcs
A random sample of about 300 images out of the full collection is more than enough to capture the style...and saves a lot of training time, preparation and caption time :)
Use a random sub-sample of 20% as a validation dataset....makes it way easier to select the right epoch.
You could also use Prodigy-ADV as the optimizer, then you don't have to care about the LR. Resolution wise, mixing 768,1024,1280,1328 yields more detail an is easily doable on the 4090.