I'm still exploring the training of Klein 9b but so far it's a very nice model to train.
I'm still playing with the quirks of the dataset but I decided to share my progress so far.
Use model with 1MP resolution ideally, it's reducing greatly the amount of limbs issue.
I will try to improve on this is the future versions.
Lora can be set .08-1 and the trigger word is a woman spreading legs.
It knows a lot about the breasts and genitals so you can be very precise.
The Lora has been trained on 9B base but can be use on both base and distilled (samples images are distilled)
Description
Beta release
FAQ
Comments (16)
whoa - are you just skilled or is klein just easier to train? This is a better nude model than most of the z-image ones I've seen so far. I thought this would be yet another troublesome flux that resists uncensored training but your examples look amazing. This will get me to finally try klein
Thanks. Easy to train but there's still a learning curve to master it and I'm still not there. It's definitely better to train on than z-image turbo because it's an undistilled base model. I expect z-image base to be the same.
Base model matters. Best base since SDXL.
Great stuff! Needs a little prompt-fu to get really good results, (like describing the color, shape and stuff of the vag) but it's a really great start. Good work!
Hello, could you share the training parameter information? I'd really like to know this, as I'm currently preparing for it and would appreciate some advice.
Hey that is an exceptionally good lora in this early stage. Could you please share the captioning tool used, where you get your dataset from, how many examples did you use and other parameters like lr and optimizer ? Also have you tried it on the 4b variant because I am looking forward to using the 4b one to train a lora. Any help is highly appreciated :)
Just started testing Klein today so can you tell me, is this t2i only or can i use this for image edit too?
Edit: Tested it! Nice work ;)
I can tell you It also works pretty well in img2img editing mode ;-)
Do LoRAs trained with 9b not work with 4b and vice versa?
Not at all.
They don't supposed to be but in one rare occasion I did accidentally selected a 9B LORA but the generation was successful on 4B model (i got lucky I guess).. nope, 4B model should use 4B LORA and 9B model should use 9B LORA
@brnfd24434343d And why the fuck did you downvote the question? Are you a fucking retard?
Just FYI with this model you can add a 2nd reference image of any [close up] vagina you want and tell it to "Use Image 2 as her vagina." Works surprisingly well.
what a time to be alive.....
BRO!!. you can use this to make "edits"
use with this > https://civitai.com/models/2324991/klein-anatomy-quality-fixer?modelVersionId=2615554 To improve anatomy.
Details
Files
spread_legs_beta1.safetensors
Mirrors
KLEIN_SpreadLegs.safetensors
spread_legs_beta1.safetensors
spread_legs_beta1.safetensors
spread_legs_beta1.safetensors
spread_legs_beta1.safetensors
spreF2k.safetensors
o_leg.safetensors
KLEIN_SpreadLegs.safetensors
Nude Legs Spreading - Klein9B.safetensors
Nude Legs Spreading - Klein9B.safetensors
spread_legs_beta1_klein.safetensors
Spread Legs vBeta - Klein9B.safetensors
Spread Legs vBeta - Klein9B.safetensors
60-spread_legs_beta.safetensors