CivArchive
    Nudify XL: Better Bodies - v1.4
    NSFW
    Preview 223144
    Preview 223185
    Preview 223184

    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:

    1. Download the file

    2. 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.

    • 0.5 weight 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

    I liked v1.3, but it often produced larger areolas than I wanted. v1.4 uses 3 totally new girls as the dataset, large and small breast sizes, same shape. As a tradeoff, the images used are of varying quality.

    Changelog:

    • Different girls used than the other versions

    • Better at producing small areolas

    • Somewhat better at nipples

    • Stronger than v1.3 - use lower weights again

    FAQ

    Comments (6)

    kwikMar 10, 2023
    CivitAI

    Hmmm... I wonder what would happen if you were to use a script to merge the different versions of this LoRA that were created with different datasets? It could be worth trying I think - and in particular if you were to train in different sizes of breasts as three different concepts.

    throwawayyy111
    Author
    Mar 10, 2023· 2 reactions

    i'm not sure about any script, but it would be possible to just remake the lora featuring the 3 new girls + the model used in v1.3 & earlier. i'll try it and let you know

    WS26Mar 31, 2023

    if can , i want have A-Cup B-Cup C D E F G ..... K-Cup ...
    SD only S. M. B. just 3 Lv. this is not good choose,And this is also very unintuitive, not as good as "Cup" for writing and understanding@throwawayyy111 

    throwawayyy111
    Author
    Apr 1, 2023

    @WS2006 probably not possible in that sense, the easiest way for it to not be confused would be to make multiple loras for each size. the ai doesn't understand really what it's seeing. a lora with all of them would probably try to blend them all

    throwawayyy111
    Author
    Mar 11, 2023· 3 reactions
    CivitAI

    The tool that I use to train new loras seems to be giving errors, including with v1.4. If you're getting this, read the updated instructions.

    throwawayyy111
    Author
    Mar 15, 2023

    this has been fixed and v1.41 has been released