A model for large breasted waifus or semi-realistic characters.
V6 Changelog 2023/06/03:
Considering this was my first and most popular LoRA, I figured it was time to improve on it. Probably the last version too. Hyperfusion is more or less the successor to this.
Less harmful to the overall style, but slightly smaller sizes overall, feels like a good balance
Wider variety of breast shapes in the dataset
Some breast shape tags were added, but hard to say how helpful they are
Uses LoRA LoCon, but does not require any additional extensions to run
Changed "large breasts" to "big breasts" to match my hyperfusion model tags
The model prefers resolutions around 640, but can do up to 768 to a degree
Training tags attached under model download section
V5 Changelog:
v5 now uses LoRA!
What changed since v4? I trained a breast size classifier to auto label the dataset which resulted in better control over breast size, and made it easier to generate larger sizes.
Note that the trigger words in v5 are ordered in the following from smallest size to largest:
breasts, large breasts, huge breasts, gigantic breasts
V4 Info:
the keyword for v4 remains hyperbreasts, if you want better control over the size use v5
Notes:
I used this to train my image tagging classifiers for breast sizes
https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification
Description
This is my first attempt at a LoRA trained model.
A1111 now has support for LoRA models!
Unlike my previous models that used the trigger word hyperbreasts this one allows you to control the size with one of the tags listed above.
If the effect is too strong just lower the strength in the LoRA extension for better quality, but at the trade off of breast size.
Training Details:
1.8k images
12 epocs
690 steps
learning rate 1e-4
only Unet training
base model Av3
clip skip 2
random flip
captions files
bucketing at 768
dim 64
FAQ
Comments (39)
what do you advise first to user of dreambooth for training
a) style
b) character
also, any ressources that helped you or is it just trail and error ?
any kind of reply is appreciated :).
I use this <https://github.com/kohya-ss/sd-scripts> But I don't have any guides to recommend, I just used trial and error. Sometimes people who post models here post their configuration too. Use that as a guide.
What is a "LORA?" Is there any special procedure for using it, or do I just put it in my models folder like a regular checkpoint?
It's somewhat similar to a hypernetwork in that it's a smaller separate set of weights that are applied to another model. But you do need the LoRA extension linked above to use it.
ok, thank you. Do you have plans to do the same with Hyperass and Hyperpreg?
If I can manage to train a better looking model than their current versions on LoRA, yea.
When following the link to download, I get this error "InvalidBucketNameThe specified bucket name is not valid."?
Civitai seems to be having issues today.
@throwawayjm Damn, okay. Thank you for letting me know, will try again later. Looking forward to using this!
This is very effective. I did notice with the LoRA it often tried to incorporate blue near the nipples, resulting in blue outfits or other coverings. Something that could be fixed with the next version of the LoRA.
I'll keep an eye out
Thank you! Reading back myself, I thought I came off a bit aggressive and unappreciative, so I cleaned it up. Still very nice.
It seems well made on a technical level but I really have to wonder about what goes on in the heads of straight men.
I can't really see how this could be sexy. It's in hilarious silly territory now.
5k downloads would say otherwise. I know I shouldn't feed the troll, but ¯\_(ツ)_/¯
Any intention to make a more photoreal take on this? Cuz that would be great
Not right now anyway. Hoping someone else will pick up that mantle. Its a bit harder to find high quality realistic morphs.
It works well, and preserves the style of the original model at low weights. At high weights, it takes over, and everything becomes kind of washed out. It works better at close ups than at a distant figure, in which it probably has some trouble identifying the breasts, and they stay the same normal size.
I'm having trouble making it create consistent sfw covered up breasts, though. Usually, it will create naked breasts on the first few steps, and if there's any "nsfw, naked, etc" in the negatives, it will then try to cover it up later, which results in it creating pasties all the time from the areolae. It would probably be helpful if the sfw images would be tagged with "covered", in case there's ever a retrain of the model. I don't know if "areola slip" is in the dataset either, but that's something I'm also having trouble with.
It would probably be easier to train a LoRA for nip slips specifically. You could do it with less than 50 images I bet.
As for strength, yea it does get harsh at high values, the model was slightly over-trained to reach larger sizes.
I'm hopeful that future text2image models will have better control over training these things.
How to install? or How to use?
If using Automatic1111, just put the lora checkpoint into ./models/Lora and then select it under the same menu as TI/Hypernetworks
Probably a stupid question but, when i select this model from the 'Show extra networks>Lora' tab, it just inserts the following into my prompt <lora:hyperbreasts_v5Lora:1>. However it still just seems to be generating the prompt based on the overarching checkpoint selected at the top. No booba. Any idea what I'm doing wrong?
That + adding one of the breast size tags should be all you need to do. Not sure otherwise, since everyone else seems to be using it.
@throwawayjm Yeah I’ve had a bit of a play around and I think I’ve gotten the hang of it. I’m incredibly new to SD so it was probably more a matter of me not having all my other settings tweaked accordingly
I'm having trouble getting the results.
Use the prompts from the images
@throwawayjm I'm trying, but not working.
@Steeltron2000 Are you using AbyssOrange as the base model, or another similar one?
@throwawayjm No.
it's using .ckpt format for a Lora. should I place it on checkpoint folder instead?
If it's a LoRA it goes in /models/Lora no matter the extension. If it's a LoRA that had been pre-applied to a model (2GB+ size) it goes in /models/stablediffusion
I'm using AbyssOrange and stuff, But I can't get it working right.
You might want to search the Automatic1111 github for similar issues. I'm not sure I can help much more. Everyone else here seems to be using it just fine.
@throwawayjm Ok thanks
It's very unreliable, I have tried it on several of my models and gives me 50 limbs monsters.
I see, I can't get the Hyper breasts.
Works perfectly fine for me. Probably don't leave it at strength 1
That is likely because many guides tell people to train on SD 1.5 - which in my opinion is a mistake since SD is the model which make these monsters.
I chose to train my hucow model as LORA based on UberrealisticDreamy_pfg - how it works out, will let you know.
Would you consider a V3 that improves upon quality and size control? Thats the biggest issues I have with the LoRA. Its a great LoRA and I use it all the time but I wish it was a little bit better now that we have models that can produce really nice looking pieces
I've been training an all-in-one LoRA that I will release soon, should be better quality and better control over size, but not just limited to breasts. And if you wanted really large sizes you can combine them both.
Details
Files
hyper_breasts_1.8k-lr1e-4-av3-LoRA_epoc12-v5.ckpt
Mirrors
hyperbreasts_v5Lora.ckpt
hyper_breasts_1.8k-lr1e-4-av3-LoRA_epoc12-v5.ckpt
hyperbreasts_v5Lora.ckpt
hyperbreasts_v5Lora.ckpt
hyper_breasts_1.8k-lr1e-4-av3-LoRA_epoc12-v5.ckpt
hyperbreasts.ckpt
hyperbreasts_v5Lora.ckpt
hyperbreasts_v5Lora.ckpt
hyprbrsts.ckpt
hyperbreasts_v5Lora.ckpt
hyperbreasts_v5Lora.ckpt
hyperbreasts_v5Lora.ckpt
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.