Shout out to Ostris and AI-Toolkit, that guy busts his hump to get us these tools fast as fuck.
What is it?
A body type LoRA, based on a tattooed, busty, natural blonde woman. All faces have been removed from training data. Tattoos unfortunately were not tagged so they are baked into this one. After reading that Ostris had solved for training on a distilled model this was the first successful training I had. This used the default settings with the v1 training helper.
.95 Strength on the right
A slim, athletic woman with a small waist, long legs, and a bright bikini, enjoying a sunny day.
Dataset
Trained on 18 captioned images at 512px
Description
FAQ
Comments (24)
Are there tattoos in the training data or did you prompt for it?. Asking because I dont like tats.
there are uncaptioned tattoos in the training data because I like tats
@DirtMcGirt k, thanks for reply
@fox23vang226 kidding aside, this was just an initial test with an existing dataset, i will caption this out in the future, possibly for a v2
@DirtMcGirt I'm looking forward to it.
Congrats!! I guess this is the first good working LORA for Z-Image Turbo.. looks good!! I am definitely gonna try this one out.. I already uploaded 4 of my Z-Image Turbo workflows (Txt2Img, Img2Img, Inpainting & Outpainting) that can use LORAs.. didn't have any good Z-Image Turbo LORAs until now thanks to you.. I am gonna post what I can come up with this here.. just one question did you train this one online or on your machine?
Thank you! Really appreciate the comment. I trained this on my personal machine, 3090 24GB in under an hour. Really impressive stuff from the z image team and ostris
Can you share your training config?
This was just an initial test using the default config in AI Trainer from Ostris
Thank you and congratulations on being the first. Kindly request to make one for proper female genitals.
We'll see what I can do :)
@DirtMcGirt Much obliged!
after 7000 iterations, i'm happy to say it was a complete failure. I'll be back when the full model is out.
@DirtMcGirt I'm so sorry to read that, and I'm sorry for your trouble but sure appreciate that you took the time and effort, 🌹
@Learning2025 of course, we're all just experimenting, thanks for the project
This lora is insane
lora key not loaded, comfy is up to date.
i've also noticed it spits out some errors to console, but it still appears to work.
@DirtMcGirt noticed my own trained lora does it, must just be a comfy issue, as you say it still works. :)
No Z Image Loras. I went to bed. I wake up. There are like 20 Z Image Loras. Amazing. Thank you and all the others <3
temporary Lora Grid: https://files.catbox.moe/nmapij.png
for me 1.4 the best
i honestly didn't think anyone even looked at those lol, thanks for the feedback
@DirtMcGirt also how it can be used like prompt word aside the trigger? can u put some example to be userfriendly use in the bio
@carlitajohn89931 This LoRA only has a one keyword for triggering it - psxay. i don't think it really does much. You can use the model strength to modify how much effect it has. I believe all the images posted by me to the gallery have full prompts you can use with the basic comfyui workflow.
here's a side-by-side of the lora applied with/without the keyword: https://civitai.com/images/114416341