Just a quick low effort model trained on PonyDiffusion to make bellies saggier.
Images were extracted from the hyperfusion dataset, so uses similar tags to that. but the only important tags are "bellysag" and the typical belly related tags.
Description
Trained on PonyDiffusion-SDXL_v6 with ~200 images.
Training Config (kohya):
lr 2e-4
frozen text encoder
batch 2
GA 4
lr scheduler cosine
dim 16
alpha 8
conv_dim 8
conv_alpha 4
tag dropout 15%
network dropout 30%
res 1024 bucket max 1024
shuffle captions
flip
ip_noise_gamma 0.02
min_snr_gamma 4
FAQ
Comments (1)
It works ok in inpaint with realistic images like throwawayjm's previous works. It unfortunately doesn't make realistic pictures from scratch even with hundreds of illistration/anime bad tags. There is another hyperpreg resource for XL that can be used together with this. But once again only in inpainting. Best shot is to use hyperpreg for the realistic base and both together to inpaint bellies after redrawing. Either way great stuff again from throwawayjm. 10/10 always enjoy your lora