This LoRA gave me way more trouble than i thought it would. And cost me over 9000 buzz to get the final product.... 😭 But here we are, so please like, donate, post or share feedback to get me more exposure so i can get more buzz to make more LoRA's 😬
See captions in sample images. Please show your creativity by posting - would love to know what you can get it to do :)
Concepts it knows and prompts/how its trained:
General Latex Clothing/Textures
Latex Hood: woman wearing a [color] latex hood head mask with [ponytail/pigtails/hair protruding from nape under mask], open eyes and open mouth style hood mask, [optional: (color) borders/trim for eyes and mouth on mask]
Optional: you can describe it as a gwendoline/gwendolyne hood but it can be hit or miss
Ponygirl: Describe in natural language...
Knows terms such as: leather head harness, leather body harness, heelless hoof boots, hoof mittens etc
Bit gag: use "bit gag in her mouth fastened to her head harness" if you want it part of it
Normal Bit Gag: The woman has a bit gag in her mouth fastened with [color] leather straps.
Bitchsuit/Frogtied Concept: (NOTE: hands and legs might break sometimes (half/single limb only instead of folded arms/legs). play with angles, strength, etc.)
Puppy Mask: She has a [color] leather snout muzzle with [color] leather animal ears.
Pose:
Kneeling on all fours: A woman is in a quadrupedal pose, kneeling on the floor with her torso oriented horizontally, leaning on her elbows and knees.
Upright kneeling: A woman is kneeling seated with her torso upright.
Restraints: Her wrists are folded to her shoulders, restrained with [(color) leather sheaths / ropes / leather belts]. Her ankles are folded to her upper thighs, restrained with [(color) leather sheaths / ropes / leather belts].
Optional: Her [hands/legs] are fully encased in [color] leather forming mitten shapes.
Armbinder: woman has her arms bound behind her back with a [color] [leather/latex] armbinder with [color] chest straps
Best views three-quarter angle, side, or rear
Straitjackets (Bolero/Regular): woman wearing a [color] leather straitjacket, [options], black leather strap running vertically from collar between her breasts, arms bound in a crossed position in front secured behind her back, [bottom style: wearing no clothes on her bottom / bottom style is white leather straps running around her crotch, she is standing in a three quarters turned pose]
Option: with black leather high collar with D-ring,
Option: bolero style straitjacket with exposed breasts, looking at viewer, she is kneeling on the floor, wearing no clothes on her bottom
Posture collar: the woman wears a wide [color] [material:leather/metal] posture collar on her neck.
Description
FAQ
Comments (6)
Thank you for all the effort. How do you train for several concepts/clothes/objects in one lora? Do you simply include several of each in the dataset and that's it?
Thank you! Getting everything included and captioned is only the start unfortunately.
It’s basically a balancing act and optimization. Have to make sure:
- the datasets/concepts have the same philosophical linkages; if you throw multiple random concepts together it may work but it may also get pulled in directions and confuse training.
- making sure your captions aren’t bleeding tokens into each other (struggled a lot with it on this one)
- the right learning rate and steps to make sure the concepts are all burning in at the same rate, while being hit hard enough but not too hard to cause overfitting issues
- the rank is the right size where it’s not too small to grasp all concepts/details but not too large to start overfitting details or mixing and matching averages
@vackyadmiral167 Thanks for the prompt reply. That's a lot of work and trial and error, but it pays off. It's much better than stacking loras, which potentially causes bleeding and degrades quality of the output.
works well with other character loras ,some concept is weaker than other single concept lora,like armbinder
Can you let me know where its breaking whether its angles or strength etc. so i can know for future fixes? In my testing it worked like 97%+ of the time and often fixed by upping strength to 1.2-1.4 without breaking much else.
works pretty well and nice variety