If you like what I do please consider supporting me on Patreon and contributing your ideas to my future projects!
This lora gives girls pubic and armpit hair which is desperately needed within SDXL 1.0. Since SDXL is so new this is not perfect but it is an improvement. The main trigger is h41ry and the ideal weight seems to be between 0.5 and 0.75. Some prompts seem to give only body horror while others you will get a high probability of nice looking images so play around with it.
Invite to my public discord: https://discord.gg/UFkPTfTDWs
Description
FAQ
Comments (17)
This is what the people want! Thank you
Looks really strange. Like it's been copied from another person and pasted into those in images.🥴
Dreamshaper, lora, adetailer, like I said its not perfect and I'll update when I can improve it
i mean thats what AI art is. just advanced automatic photoshop.
Overcooked effect. Properly cooked will blend nicely with the main image.
Looks good enough already!
Yeah still a long ways to go before it's better than 1.5 but we'll get there
@getphat I think what makes SDXL less 'usable' than 1.5 is not only the VRAM usage but also the fact that it is not trainable on Colab for free, this is a huge problem if you want to try out things before actually working on your local environment... Looks like right now there is no way to use SDXL in any way for free online, in private i mean.
@gsgsdg True, and in regards to this model specifically, it still doesnt like more complex prompts and poses, but I will keep playing around with it to see if I can improve it
@getphat I have the feeling that this, like any future lora, have or will have problems with complex prompts because Stability AI trained the text encoder on the generated images that users found to be the best on Discord, this technically makes some sort of shortcut for better quality but also ruins prompts that were not 'thought of'
@gsgsdg Thats interesting, I'll take this into account when adjusting future training parameters, right now I'm retraining on prodigy optimizer with a text encoder and unet learning rate of 1, so I'll see if that makes any difference
@getphat I'm not entirely sure that was their workflow in fine-tuning, but it's possible because i managed to do the same (kinda...) with SD 1.5, i suppose it's some sort of reinforced learning that is a little behind to the methods of Midjourney, but right now I'm almost sure that Midjourney V 5 is equivalent to SD 2.X as for text encoder
wow, sdxl does pubic hair way better. wish I could run it. lol
goshdarn 870 MB :DDD
I gave it a shot and really looks nice.
Only thing I noted that the default value is higher than recommended max in the Online UI.
I tried it on SD XXXL and it worked ok for pubes - gave me lots of tan lines, though. I didn't have much luck with armpits.
For my experience, that works fine, but highly disforms faces compared to SDXL_hairy_v1.5_LITE