breastInClass -> nudify XL.
Are you enjoying fine breasts and perverting the life work of science researchers? Consider donating a dollar to the cause. Anything is appreciated :)
tl;dr:
Nude female frontals. Shapely breasts with small nipples. Strips away layers of embedded censorship in models.
⬇️ read the instructions ⬇️
What it is
This is a LORA that I use in text2img to get better NUDE FRONTALS.
It alters the body shape, particularly the breasts.
It's best at weights under <1 with accompanying keywords.
The data set is a collection of fully nude adult women.
What it is not
The dataset is all naked women with no faces. Any effect on clothing, men, flat chests, or faces is unintended. Funny, but unintended.
While appreciated for the exposure - using this for fully clothed women defeats the purpose.
It only works when you include it in the prompt (saddened I must tell you this).
It does not bias towards asians. I just like generating them. There are no dark skinned women in the data set unfortunately, but I'm not averse to making a version for this.
Installation:
Download the file
Place it in models\Lora folder (webui)
Usage:
You can also add it to the prompt by clicking "lora" tab in automatic1111:
You can also type it into the prompt, but it will not work if you make a single typo.
ex <lora:nudify:0.5>
The XL model requires the XL base / XL base with VAE model. It doesn't seem to work with the refiner, or anything else.
Adjust the weights. 0.1 - 1.0
Tips:
This is a fairly strong LORA best used in conjunction with low weights and other keywords like "naked", "topless". Work to get your image close to NSFW first, then add this to push it over the line.
0.5weight is pretty good for most purposes.Very high weights can cause the face to crop out.
Adding "small breast" or "huge breast" keywords can add more size variation, keeping the shape.
Adding stubborn loras or conflicting terms (e.g. "shirt") to the prompt can burn the image.
Base SD models are stubborn and don't like to be made NSFW. Expect clothes, blurred nipples. Try different weights and add more keywords!
Description
The tool I used to make the lora fixed the issue that would print an error on first run. The data set, steps, configuration is all the same. There should be no/minimal difference in the generated image.
CL:
No longer produces an "invalid keys" error
FAQ
Comments (81)
Works great, but very strong impact on the image
Can turn stylized to realistic and crop out head … use carefully
Yes, says this in the instructions
"Couldn't find Lora with name breastinclassbetter_v141" - what do I do wrong? called <lora:breastinclassbetter_v141:0.6>
Did you try adding it from the pink icon under generate? Might be a file name issue. It’s case sensitive.
Check the “usage” section above
For any Lora (like embeddings) the name of the Lora in your prompt is simply the name of the file.
If you rename it as BetterBreast for example, then clicking on the Lora icon will add to the prompt :
<lora:BetterBreast:1.0>
⬇ ⬇搭配这个破万人使用的模型效果很好！日本/韩国论坛目前最推荐使用的！大家可以试试！⬇ ⬇
https://civitai.com/models/16599/beautypromix
有可以分享的论坛吗
又可以共享的论坛吗
有可以共享的论坛吗
你能分享一下论坛吗
有没有可能分享的论坛？
I think those looked like plastic implant, not natural
all girls in the data set are natural. some girls just have perfect titties ¯\_(ツ)_/¯
So this is only for nudity? can it for make better breasts even using clothing?
which VAE to use,840000 is very terrible
That should have nothing to do with this Lora
clearvae\anything3.0
is this just for female models?
Yes
Does this LoRA influence the face? I've already noticed some changes in faces because of this LoRA, but I think it might be an indirect change.
Different from the description, this LoRA influences the breasts even with clothes, for example some white shirts tend to be transparent or simply the nipples are marking the clothes.
These would both be unintended effects. There are only naked bodies with a cropped face in the data set.
@throwawayyy111 I read somewhere on Reddit that breasts are the age and appearance of the face or maybe the entire body. I understand when you say it's not intentional.
Big breasts look too hentai-ish, shiny. It'd be great if you made a version without that digital painting looking breasts
i think the base model would be better for this
When I generate why no face appear
this is in the instructions
lower the weight. This can be done by adjusting the number at the end of the triangle brackets
Just curious, for your training data, how did you go about tagging? E.g. did you use descriptors for breasts / body shape, and did you use BLIP or some alternative tagging?
Building my own body LoRa and would love some tips!
I used BLIP but the results were not very accurate. I went through the data set by hand. All images were tagged either "large breasts" or "small breasts"
Does anyone know of a similar model that works well on anime?
Trigger Words was not right, should be <lora:breastinClass:.4> is
should be fixed
无法加载Lora怎么办
I am currently working on releasing the XL version. It's MUCH more accurate and covers both small and large sizes. Hard not to release in its current state, but needs more testing.
(Related to todays XL_v1 release)
The LORA doesn't show up in the appropriate tab of Automatic1111.
Adding the suggested <lora:nudify:1> to the prompt directly (incase the UI in the foreground might be faulty but it still registers in the background) returns a whole list of errors.
Are you using the SDXL v1 w/ 0.9 VAE model? I saw this while trying to use the refiner model
Let me know if this solves your issue, I'll add it to the instructions
@throwawayyy111 Using the SDXLv1.0 model actually made it work, yes.👍
Up until now I was under the impression that LORAs can be applied to any model regardless😅
Why is the file size so big?
can you make it smaller on the next update?
because it's sdxl and no
@throwawayyy111 What rank did you use to train the Lora? That influences the file size. You could probably get away with lowering the rank and maintain good performance.
use kohya_ss to resize the rank down to 64 or 32....
Lora won't show up in my Stable Diffusion on Windows OS. I put the files in the Lora Folder. How can I fix this?
are you using sdxl
@throwawayyy111 I don't know I'm using this one. https://github.com/Stability-AI/stablediffusion with the sdXL_v10VAEFix checkpoint. Is this the correct version I need here - https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 ???
I have the same problem.
Using https://github.com/AUTOMATIC1111/stable-diffusion-webui
lora won't show up, I already put the vae, but I don't know if I put the right one in the right place
win10
same, wont show up
same here
I found out why. XL Lora's wont show up if you're not on an XL checkpoint. Switch your checkpoint and the lora will be in your lora's dropdown.
@Novellus Hi, would tell me how to switch the checkpoint, or any reference? Thx very much
got it! In my SDXL it also doesn't show up at beginning, finally found in SDXL's settings, under 'Extend models' term, check the 'lora show all', done.
Nice work, but every body is ribcage skinny.
yeah - wondering if this is an sdxl thing. sometimes makes them have very long torsos. i'll try to compensate for it with the next version
Definitely would like to see you add some darker skinned beauties to your training. Native Americans, Hispanic, Indian, Arabic, African American... those ladies have gorgeous skin.
Going from 1.5 to XL just doesnt feel right. i honestly think you should've made new page for this one
the old downloads are right there
can you please tell me which version is <breastinClass: d62716438a56>?
v1.41
Hi, are you the one previously have the "breastinClass: d62716438a56"? That LoRa is quite good, please consider to bring it back, thanks !
Yes, I am the author, and this is the same LoRA. Look into the older version downloads
Ah I see, I though it was a different one cause u changed the name. Thanks!
RuntimeError: mat1 and mat2 shapes cannot be multiplied
Anyone getting this?
I was getting this when trying to use controlnet. Are you?
@bounceboi are you by any chance mixing resources for sd1.5 with sdxl resources?
@throwawayyy111 same here. @eurotaku without mixing 1.5 and xl, just xl checkpoint and controlnet. got any hint?
@4tyzap check if the controlnet model you use is designed for sdxl
@eurotaku thanks, noted, resolved for a single controlnet model. (off-topic) I can't use together this https://civitai.com/models/136070?modelVersionId=155414 and this https://civitai.com/models/136070?modelVersionId=154923 as controlnet 0 and controlnet 1 simultaneously. if I divide those to different steps (for example controlnet 0 for 0...0.25 and controlnet 1 for 0.26...1.0) then everything goes well. any remedies for the case?
@4tyzap hard to say, can you join our discord and post an error log in the support section?
Am I the only one to use it with negative weight in order to get more natural and more realistic bodies ?
Good man!
Как поставить отрицательный вес?
@228afrin228 <lora:nudify-000005:-0.2>
This LoRA won't run on a GPU with 8 GB VRAM - SDXL checkpoints are 6.5 GB and this LoRA is over 1.5 GB, pushing it over the limit. Some tools crash, others revert to CPU rendering. ☹️
should have a mini version up tonight but I expect that the quality will vary
lite version is up.
@throwawayyy111 Thank you! 😘 I will give it a try.
@throwawayyy111 Should I always use non-lite v1 if I have the disk space? Or are there any cases where lite produces better results?