Illustrious will recognize this prompt. This LoRA will help make it more complete.
Main tag : slingshot swimsuit
Secondary tag : wedgie, pubic hair, partially visible vulva, pussy peek,
Extra Tag : clothing aside,
Color tag : color slingshot swimsuit,
LoRA str : 1
Training Pony LoRA Powershell info :
resolution = 768
Batch size = 3
network_dim = 12
network alpha = 12
mix/save_precision = BF16
optimizer_type = Radam
optimizer_args_other = weight_decay = 0.1
GPU RTX 4060Ti 16GB
Torch 2.7+cu128 + Xformers + gradient_checkpointing
Description
LoRA dim = 64
slightly decreased precision
reduce file size
tag works correctly
FAQ
Comments (12)
Will this work with realistic generations?
Also, is the base model really SD 1.4, and if so, why? 🤷
Base model means Traninig model ,you shouldnt care
@ailand I know what it means. I am asking why.
@klaes It's not a choice so choose SD 1.4 instead
Why not have two separated LORAs so that the slingshot can work more consistently, because personally I only care about the slingshot bikini? Also, why are there three versions, what's the differences?
- dim 32-64-128 to minimize space usage
- As for the two swimsuits they are clearly different
Therefore data can be added to train LoRA together
- there is no difference between the 3 versions dim 32 requires str in the prompt to work properly [example : (slingshot swimsuit:1.2)]
Maybe the trigger word should be changed for the eyepatch bikini because it mixes with actual eyepatches and they can be hard to remove with negative prompts.
can you make this as well? I've tried training a LoRA out of it but to no avail.
i don't know what the name of the outfit is
OP doesn't know the name of the outfit and I can't see the pic because I can't remember my Pixiv log in, so recommend asking ChatGPT what the name of the outfit is called by describing the outfit in detail. That usually works for me. phrase it like, "what is the name of the outfit that [insert details here]"
It said V dress, probably added revealing clothes.
This is epic, thanks so much. I'd been trying to prompt for these without any luck. Seems to work well on realistic models also. Using it with Realistic Vision and am cranking it to 1.5 and getting nice results. Not always perfect, but much sexier than what I was getting before.