So this is a test its not amazing or overly good, the idea was just to see if I could take a bunch of SDXL renders, using wildcards for variation and then train flux on that to get something half decent.
I uploaded 3 different steps counts the higher you go the more consistent but the more burnt it the data becomes making it less flexible. Yes its not the best but I figured everyones still testing and playing these where only captioned "sucking a penis" with some images of penis's in the data set. feel free to change the captions, re-train whatever.
While not essential I would appreciate a shout out if you used my dataset to make something better xD. not sure if I will re-train this it was purely a hypothetical test. Ideally using a tool to remove facial likeness and getting an equal amount of faces would remove the same face syndrome but it also might just be my prompting anyway go nuts.
Description
FAQ
Comments (2)
Thanks for sharing your work and that explanation.
Would be cool to see your training data set, so one could learn an derive from that. (interested in the captioning, size of data set and quality).
The training data is in the files part of the downloads for people to play with. I have heard some people can get better results using longer captions or wording things different ways, this merely turned out decent so I thought I would share :)