My first attempt at training Z-Image on a 3060. It feels almost as easy as training SD 1.5 used to be! Will definitely try to make more LoRAs -- open to requests.
It can have trouble with more than three girls -- you might have to try generating a few times. You can get better and more consistent results by being very deliberate and precise with your prompt for each girl.
Upcoming V1.3:
Differentiating between sucking, swallowing, and "throating" a penis.
Changed "making eye contact" prompt (which Z-Image has trouble with) to "staring at camera".
V1.2 Changelog:
Added some images of just a POV view penis to training set
Went back to older, harsher training settings -- but using only a partially trained output for better flexibility.
V1.1 Changelog:
Fixed typos in some captions (yes, it's all manually captioned)
Added more asians and latinas into the dataset
Set rank to 16 from 32
Set learning rate to 0.0003 from 0.0004
Description
FAQ
Comments (8)
awesome lora sir. now i'm gonna wait for t2v wan :)
The version for Z-image seems to have a very strong style, it overrides my trained character loras to the point their races changed from asian to caucasian/latin
Try lowering the strength and specifying that they are asian in the prompt.
@degeneration_zee Did you train on the de-turbo model or the normal distil?
@tenstrip I trained it on the deturbo model, is it worse?
@degeneration_zee Should be better. I think really the lora game will start when the real model weights come out.
I added more asians to the dataset and lowered the training rate. I also recaptioned all the white girls from "girl" to "caucasian girl". Give it a try, let me know if it's better. :)
@degeneration_zee Yeah now it gets much closer to the intended character from my lora!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.