V3
New and improved restructured training and captioning, higher quality image training, trained in nudity so now works without additional nsfw lora much better, keeps style really good
Lora strength 0.4-1.25 what i tested
No trigger, most prevalent phrase in captioning - leaked nudes style
Better to use either way with already finetuned models, but does work for any Flux Klein 9b, if there are errors, please leave a detailed comment, or hit me up in messages, i will really try to help and understand the issue myself
Example prompts are in images and more phrases that are semi-consistent - casual, intimate, selfie
V2
Additional training on top of base lora with new smaller set of images/poses
Overall quality update, increased amateurishness of results, style adherence
Lora strength 0.2-1.2, needs either already finetuned on nudity model or nsfw lora in addition, lower strength for 9b with turbo and img2img
V1
Trained on Flux Klein 9b Base
The dataset consists of images from free resources posted with consent, like reddit, onlyfans etc...
Style adherence of amateurish self photo/selfie is pretty good
Works well on both Klein 9b and 9b Base, works well on basic model with a little prompt push or/and with nsfw loras, as well as already finetuned models, change prompt/strength for better results. Amateurishness of results vary from these different methods
strength 0.75-1.25
If you want some improvements i am open to suggestions
Description
Overall quality update on top of original version, better style adherence
FAQ
Comments (11)
Is it possible to have a breasts or pussy mod looking down or a nude version looking up pussy focus from the bottom? :)
I mean, I will try, thanks for suggestion
the v2 lora ruins gens on wan2gp (not using flux base), they end up all ghostly and weird, haven't tested 1 yet
From my Knowledge v1 probably won't help, my guess it's quantization+mismatched lora weights.
I will try to train LoKr(Lycoris) next with a different tool, the wan2gp was designed around that support, and it's probably overall more supported on different flux models
yes .. .sry, v2 is bluring/softenign all
@OlovoDude it might be wan2gp or not, i tested lots of loras and all of them work good, the ones that failed were made for flux9b base, those show noticeable noise from lack of steps
on Non Base 4 step getting body horror.
Does not work on f2Klein 9b. Ends up looking doughy and body horror, as another person commented.
i mean, there are literally my images using flux klein 9b with 6 steps on Lora showcase
also my workflow used is also posted
Using some other way than comfyui it can maybe only accept LoKr type and not Lora for Flux, that's my only guess, since there are no such comments on my LoKr for Flux
@OlovoDude No, I'm using comfyui and correctly, so im going to hazard a guess and say its the esoteric workflow you use that allows this, rather than user error on my and the other guys part.
I will be glad to post the doughy pictures to the page, if you like. I was not doing it out of politeness.
@badhandproductions sorry, maybe i was coming off rude, didn't mean to offend you, thanks for more detailed clarification, i looked at your images and tried using the same model you point out, and with just my lora enabled it does produce body horror even on my workflow, i tested it different ways + on workflow from the same creator who uploaded fp8 on civitai, and it semi-works with that model and nsfw lora 1.0+my lora 0.2-0.3, my next guess is key mismatch, because it's not Lokr, it's either way more compatible with any modifications of flux, as i see + basic fp8 maybe is somehow more censored, because it produces nice results even with my lora enabled but with prompting not mentioning leaked nudes, nudity or other flagged texts. My new Lora however works fine on this model out of the box, but it was just more trained on nudity, so there is that