Makes breasts bigger.
(IMPORTANT: In my opinion V1.0 is better than v2.x. I'm providing the 2.x file only so you can try it out if you like. Combining both versions in the same prompt will get you even bigger breasts, but 2.x causes a higher rate of trash images.)
If applied to a picture of a woman, will tend to increase the size of her breasts. Also skews images towards realism.
This will work somewhat without any specific keywords, for example:
vs
However, it responds very strongly to any mention of breasts, especially size modifiers such as "large", "huge", or "enormous":
vs
The above examples are generated with the Deliberate model based on SD1.5.
Strongest effect seems to be with the exact phrase "enormous soft natural breasts":
vs
Above example generated with the UberRealisticPornMerge model.
Other key phrases that get good results:
Suggestions of "stretching" clothes and "bursting" or "bulging" out of clothes.
"Detailed realistic skin texture" or variants on that phrase.
"Soft" and "Natural"
Problems that may be fixed in future versions:
Way too eager to get girls naked. Will straight up ignore "NSFW" and "naked" negative prompts a lot of the time, or make clothes partial or transparent. Use increased strength on negative prompts to avoid this.
Is trained to understand prompts related to freckles, veins, areolas and nipple size, but comes on very strong at normal strength. Dial it down to 0.5 or less to get better results.
Duplicating synonyms of "breasts" in prompts (e.g. "tits", "boobs") will tend to get additional breasts on the same woman instead of larger ones. If you're into that, go for it.
Does not do well at complex poses. Though sometimes will actually fix weirdness generated by the base model. For example, seems to improve Deliberate's attempts at laying down poses.
Please review if you like this! Comment if you find something that doesn't work, or if you've got a use case that needs improving for the next version. Show me your images!
Additional interesting notes: this hypernetwork has unexpected interesting effects on images not of women. Landscapes just seem to look nicer? Scenes seem better composed with more interesting backgrounds? It especially seems to like water and trees, but ground textures seem to be affected positively too.
Description
As an experiment, I re-trained the hypernetwork from scratch using the same image set with much more meticulous tags. The results were not amazing.
The positive:
Now respects NSFW negative prompt much more reliably
About 50% of the time will generate larger breasts than version 1.0
Combining both versions in the same prompt can create some truly gigantic breasts
The negative:
Much more likely to generate deformed horrors. Especially extra breasts, double-heads, extra hands, and creased inverted nipples.
Weirdly less realistic natural breasts. Much more of an airbrushed plastic look in this version
I'm currently re-training again with the results of what I've learned, so hopefully an improved version coming in the next couple of days.
FAQ
Comments (7)
I've messed with this some, and it's not quite what I was expecting. It doesn't seem to make things bigger necessarily, but just guarantee that they are large instead of it being uncertain with regular URPM. That model already does a fairly good job of making large sizes, though sometimes it doesn't do what you want. This hypernetwork, however, seems to fix that one problem but add more. It tends to make weird streaks or creases or lumps and occasionally it generates some horrifying Cronenberg thing. That said, it can also generate some very nice images. Also, I would appreciate if there were triggers for different sizes, as right now I can't really make it differentiate between ones that are as big as watermelons and ones that outsize automobiles.
Agree all around. Which version of the file are you using? I find the creases are a problem of 2.x which I consider a failed (but still interesting) experiment. If you're having the same problem with 1.0, do you have any image examples (with prompts & seed) in the metadata that I could run some test on with the version 2.0 I'm currently training?
I didn't realize I was still using version 2 when I made my original gallery post. I edited it to include two images generated with version 1. In the first one you can see some small streakage radiating from the center, while in the second you can see a separate issue of the hypernetwork sometimes adding text in the corner for some reason.
Why do using the word breasts mean you don't have a bra?
That's something I'm working on for future versions. Mostly it's because the training images were mostly clothed, naked, or in bikinis. v1.0 isn't great at bras.
For now, you can usually handle it by editing prompts. Add a description of the bra in your prompt, or add "NSFW" or "exposed breasts" to your negative prompt.
works quite well, despite me being new to stable diffusion. Using it with FantasticChix and vae-ft-mse-840000-ema-pruned has given me some stellar results. Though you're right, it loves exposing the skin, though in some instances ive been able to have my generations fully clothed.
Great model... ! Just wondering why there is no breast model or checkpoint where the nipples have been trained as well... Weird since that is the crown of the jewel... ;D Just meant as an inspiration for later versions...