Krea 2 v1.1: I thought the outputs were a bit too Flux 1.D-level / plastic-y so I re-washed the dataset with my current Flux 2 Klein enhance/upscale workflow to help with that. Looks a bit more realistic and less Pony/Flux 1.D quality level outputs. Also should stack well with character LoRAs (examples provided).
Krea 2 v1: Re-trained w/ Krea 2 with updated caption set. Should be more performative with prompt adherence and whatnot due to it being Krea 2.
F2K9B v1: Pretty much the same dataset of ZIT v2 but images were treated/upscaled to much higher quality with realistic skin and more varied color grading with a Flux 2 9B + Ultra Real Edit Mode-wash . Results are much better realistic outputs as far as skin quality goes.
Image gallery examples contain metadata in case you want to follow along on exact ComfyUI generation parameters but the nutshell are the standard ones - BFL Flux.2 9B base, Euler, 4-step, and no other stacked LoRAs. If you want to use increased step counts, adult oriented merge checkpoints, stacked NSFW / Detailer LoRas and/or more advanced samplers then your results will vary (probably for the better).
... and yes, it can do hetero interactions as well like the previous version for ZIT. Just prompt accordingly, you breeder. 😆
ZIT v2:
Update: Turns out this LoRA is a damn pansexual and can create hetero content, if so prompted. I guess I should be supportive of its digital life choices despite my sus dataset training. Deep down I suspect this is my husband's doing with his sportsball watching and whatnot - that and its ZIT genetics.
--
I created a LoRA to assist creating scenes of nutsack worshiping since Z-Image (and Flux) generally will create scenes of mouth-on-penis despite specific prompting. I've noticed that of the checkpoints I regularly use, you might get the concept eventually with the right seed (heh), but usually requires the use of CN or a second pass inpainting run.
Well now your boys can get that lovin' attention they so deserve with this little number.
The dataset includes about 440 HQ images to get a good diversity during the training process. I used a large dataset because of ZIT limitations on using multiple LoRAs, I made sure there was a varied viewing perspective of peens of all types (uncut, cut, big, small, black, white, with or without cock ring, thick pubes, no pubes, etc.). So, with Z-Image's strong prompt adherence in general, using this you can specify things like licking, sucking, smelling, covering face / eyes / nose, on forehead / head, etc., and get good results of the full .. package. One note is uncut peens might take a few more generations if that's what you're going for. If I create a new version I'll correct that imbalance in the dataset.
It shouldn't need a trigger word but it can help, and even short prompting can generally yield good results so long as you give a little direction.
Images in gallery are all txt2img generations without img2img 2nd pass corrections.
I recommend strength about 0.8 for a more realistic skin texture.