Trained on deliberate_v2. Training images are taken from high res pinup paintings posted on Heritage Auctions. All of the subjects are female.
It gives a nice painterly style to the image. It will also give the ladies the vintage hairstyle (you may or may not like that). This LORA can easily do exposed breasts. Sometimes it makes weird dresses and skirts because the training data features scenes where womens' skirts are being blown up to expose their underwear. I may try to compensate for this in a future version.
Description
Trained on deliberate_v2
*Seems to have abnormal difficulty with hands and feet. I'll try to address this in a future update.
FAQ
Comments (9)
Works great! Thank you!
Great Lora, very easy to get that retro pinup effect. I was REALLY struggling with the hands tho. Almost all of mine were coming out with mutated carrot fingers. I think because of the style, bad-hands-5.pt does not help much.
Me too, I'll see if I can address this in the next update. I wonder if it's because of the nail polish.
The update will have addition training images. I might add more 512x512 closeups that include the hands and that may help.
Am I the only one who has had problems with this lora? my stable deffusion does not detect it. This error pops up: NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.
Any idea what to do?
Edit your webui-user.bat file. The command line args line would look like this:
set COMMANDLINE_ARGS=--no-half-vae
Nice work! Which model was used for training?
Thanks. It was trained on Deliberate v2.
Has anyone tried using it in Fooocus? Since it's not an SDXL LoRa I assume it wouldn't work, would it?
I haven't tried FOOCUS yet. My graphics card is only able to handle training SD1.5 but once Kohya starts supporting Stable Cascade LoRA training then I'll make versions for my LoRAs.