This lora provides perkier, pointier breasts with larger, puffier nipples... without changing anything else.
Minimal changes to styles, lighting, color, people, backgrounds, and everything that isn't breasts.
Maximum interaction with hair, clothing, plastic, liquid, paint, glitter, materials, objects, other loras.
TRIGGER WORD (not required, flexible wording, sometimes better without): "small pointy breasts with large puffy nipples" was the main phrase included in training captions.
Note: mentioning breasts/nipples in prompt is less about "activating" the lora, and more about creating favorable conditions for the lora to work.
STRENGTH: Optimized for 1.0 but continuously adjustable up to limit
Note: I challenged myself to create all samples with 1.0 weight, but I definitely wouldn't suggest this, your life will be much easier if you adjust up or down as needed. It's a pure smooth change from zero up. The ceiling where it starts to break down depends on many factors.
IMPORTANT: It's not a magic wand, it's highly dependent on the generation beneath the lora. It doesn't "draw breasts", it only changes the shape of existing breasts. It doesn't make klein better at anything it couldn't already do. If you get weird results, turn the lora off and check if klein is doing something weird underneath it. Occasionally klein does weird things if "pointy" and/or "puffy" are in prompt. A vast majority of problems can be fixed by tweaking prompt, seed, or lora weight.
EXCEPTION: It does help remove "kleinderwear", the phantom underwear klein adds when you ask for nude. But for best results find a prompt/seed without kleinderwear.
OTHER TIPS:
See sample images, all samples use basic settings (klein-9b-fp8, euler/simple/4step, 1.0 lora weight) and all prompts and seeds are listed (if not shown, seed is 0). Or just drag the images directly into ComfyUI. Some of the samples are easy to replicate, some are extremely difficult.
If all you care about is nudes, no tips really needed, just convince klein to give you a topless woman and turn it on, adjust weight however you like.
For more complicated stuff, try emphasizing nude first, then add clothing/material. Use phrasing like "completely naked, covered in..." or "fully nude, wearing only..."
Try mentioning that the main focus of the photo is the woman's breasts.
If nipples show through when they shouldn't, try lowering weight or different seed or not mentioning nipples in prompt.
If nipples don't show through enough when they should, try raising weight or different seed or more aggressively mentioning nipples.
Cartoon/video game styles, puffy/pointy is more frequently misinterpreted, and klein often exaggerates breast size and roundness so higher weight is needed.
It doesn't give more control over breast size than klein has by default. You can change size with prompting, but use with a breast slider for more control.
It doesn't make klein understand words it didn't already know, or generate things it wasn't already willing to generate. Use with other NSFW loras for more explicit content.
Description
v1.1
Corrected precision loss due to scaling error
Cut rank (and file size) in half
Slightly less powerful (~5%) at 1.0 weight
FAQ
Comments (7)
This works really well!
Any tips on training you can share to have it affect only the part you want to train?
Which trainer did you use?
Thanks! I used ai-toolkit. I think most important is just curating a quality dataset and spending the time to create good captions, then masking what you're targeting and only training specific blocks help a lot to isolate. But yeah a ton went into this project I will see if I can write an article about it sometime when I get a chance. Only been doing this a few weeks though so I barely know what I'm talking about yet
"only training specific blocks" would love to learn more about how you identified the blocks to target and how you did this. I think there's a config only setting in ai-toolkit to mask blocks?
@nonamezoname7621 I trained a lora with all blocks, then used the "LoRA Loader (Block Weight)" node from Inspire Pack and basically just tried flipping everything on and off in different combinations and comparing images, narrowed it down to a few obviously important blocks, they were grouped so I assumed there was some kind of continuity going on so kept a couple extra on either side just in case they were important for subtle details or edge cases. And yeah in ai-toolkit config under network_kwargs you can do only_if_contains or ignore_if_contains to specify which blocks to train
@vaaaaaan So with masking, is it just making everything around the target white/black on the training images? or is it providing a black and white mask accompanying the training images, like how they talk about it here for Onetrainer? https://www.reddit.com/r/StableDiffusion/comments/1g1u3ng/is_there_any_well_explained_tutorial_of_how_to/
@Kamk accompanying black and white mask images, same name as training images in different folder, target = white. I used the mask editor in comfyui to do them manually for this project, can also use something like Segment Anything depending on how much control you need
@vaaaaaan Appreciated! Now to do some testing 😋














