Breast Size & Shape
Trained to convergence on 332 images with Natural Language and captioning that has been checked by hand and edited on 70% of the BLIP-2 Captions. (Version 2 is Dreambooth)
WD-14 tagging is a custom dictionary with manual appending.
Dreambooth Version
Avoid using "nude" if possible.
Use negative -"Large Breasts" when using positive +"small breasts" or +flat chest
Use Natural Language with VIT tagging only where needed.
Use (Dreambooth) at 1.0 for detail 2.0 lora value for Shape/Size changes.
Saggy Breasts Version
Use at or around 1.0
Best use at 512x768 (Up scaling tested up to 768x1280)
Note that doing Image to Image checkpoints will "correct" the nipple position
Beta Version
Use with trigger at or around 1.0 Lora value
Description
FAQ
Comments (7)
I see somewhat of a taxonomy as the set of trigger words, but no representation for the itty bitty titty committee. Did you only train on big boobs?
Flat chest and small breast have both been trained (At a higher rate do to them being outweighed...no way to express that without being a pun). Small detailed differences in a flat chest or A-Cup could be described an trained, but likely would need to be separate from a universal tool (Or well described during a finetune where the weighting is compared full model)
@Felldude Oops, overlooked "flat chest". "Small breasts" isn't on the list though. Could you provide your whole taxonomy?
@unclehater I listed the highly trained clip terms, you can also use natural language as BLIP 2 uses tits and boobies sometimes. Small breasts, large breasts, medium breasts, flat chest are WD-14-Vit standard terms but have a high error rate (I manually set them for this lora but the model accuracy will still be poor)
I created an article with lots of examples I generated from this lora.
Don't know if you feel like testing again but I did a rework of the set trained in Dreambooth
Here's the v2 review https://civitai.com/articles/6125