facialized
Details
Download Files
About this version
Model description
facialized: a new LoRA for facials on women
LoRA to generate photo-realistic images of women with cum on their face.
Include in your prompt <lora:facialized:1>, cum, facial.You might want to include in your negative prompt cum on breasts, cum on body.
The model works best with low steps and CFG, I get good results with 10 steps and a CFG of 3 or 4.
I am open to advice/help to improve this LoRA. If you are willing to help, you can send me a mail at the mail address at the end of this description.
Training methodology below.
Know issues with the model
As rightly noted by some comments, the model is still imperfect. Here is a list of a few issues I found, along with ways to circumvent them when possible.
I will share tags present in the training dataset that might help you circumvent some of these issues in the format the tag in question [number of images out of 2676 that were associated with this tag]. For example smiling at camera [486] means that 486 images in the dataset (composed of 2676 images in total) had the tag smiling at camera.
Cum on the body / torso
Most of the generated images depict women with cum on their body / torso / breasts. It is currently hard to remove it and ensure that cum only appears on the face.
Here are some tags present in the training dataset that might help you with this issue:
cum on breasts [195]cum on body [472]cum on clothes [10]
Bad eyes
The model might generate faces with non-symmetrical eyes or ill-formed eyes. You can try to alleviate this by using tags such as:
symmetric eyessame color eyesbad eyes(negative prompt)strange eyes(negative prompt)
These tags do not appear in the training data but help the underlying stable diffusion checkpoint generating correct faces.
Training methodology explained
Disclaimer: this is my first time making a LoRA and I am more than open to advice to improve it! I am detailing my methodology below, if you have any idea on how to improve it please feel free to comment.
Dataset
The full dataset is composed of 2676 images, hand-picked from one source. Their quality varies greatly, but nearly all of them show a unique women with cum on her face. Some outliers show 2 women (~10 images).
The image sizes are very disparate, I am reproducing here a count of the number of images for each resolution that has 10 or more images (there are 1454 different resolutions in the whole dataset).
154 3024x4032
98 1536x2048
96 2316x3088
51 960x1280
46 750x1000
36 768x1024
27 853x1280
25 510x680
25 1920x1080
22 1080x1920
20 1000x1333
19 1280x960
17 1024x768
14 1280x1707
14 1000x750
13 854x1280
13 2268x4032
13 1200x1600
12 2448x3264
11 1280x1920
11 1067x1600
11 1024x1365
10 600x800
10 2000x2666
Filtering and tagging
I used https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Dataset_Maker.ipynb to filter the dataset. Duplicates have been found with FiftyOne AI and a similarity threshold of 0.985.
Image tagging has been performed locally in stable-diffusion-ui stable-diffusion-webui-wd14-tagger extension. Duplicate tags are removed, I used the wd14-vit-v2-git interrogator with a threshold of 0.35. I also added the additional tags "facial", "cum", "1girl", "1women", "face", "sperm".
Below is a list of all the tags appearing on more than 20 images in the dataset. Note that these tags are mostly obtained from an auto-tagging procedure (as described above).
1girl 2676
1women 2676
cum 2676
face 2676
facial 2676
sperm 2676
realistic 2631
lips 2192
solo 2160
looking at viewer 1459
breasts 1262
long hair 1085
brown hair 1031
nude 1011
black hair 997
smile 946
nipples 885
jewelry 883
closed eyes 804
blonde hair 718
brown eyes 706
open mouth 663
freckles 592
1boy 554
teeth 539
hetero 513
solo focus 496
medium breasts 496
tongue 494
earrings 486
penis 479
cum on body 472
necklace 449
upper body 334
tongue out 328
uncensored 328
small breasts 325
nose 298
indoors 290
large breasts 279
blue eyes 262
grin 255
mole 253
blurry 253
short hair 246
cum on hair 242
piercing 234
cum in mouth 228
from above 227
cleavage 224
closed mouth 212
tattoo 209
makeup 201
cum on breasts 195
forehead 185
underwear 177
shirt 171
portrait 171
sitting 165
erection 163
glasses 141
navel 138
looking up 136
black eyes 127
lying 126
parted lips 124
bra 119
pov 119
oral 118
fellatio 111
kneeling 110
bed 106
male focus 103
blurry background 96
multiple boys 95
mole on breast 89
handjob 86
on back 84
topless 84
ring 83
bukkake 82
nail polish 81
green eyes 78
pubic hair 75
choker 73
one eye closed 73
collar 73
nipple piercing 73
photo background 68
2boys 67
eyelashes 64
hoop earrings 63
male pubic hair 62
dark skin 59
cum on tongue 57
multiple penises 56
downblouse 56
close-up 55
thighhighs 55
what 55
twintails 53
panties 52
outdoors 52
testicles 52
multicolored hair 52
censored 52
flat chest 49
red hair 48
barefoot 48
pillow 48
tank top 47
bracelet 46
group sex 44
selfie 44
wet 43
bare shoulders 43
tears 43
clothes lift 42
collarbone 41
watermark 41
completely nude 41
bdsm 39
lingerie 38
half-closed eyes 37
bangs 37
shirt lift 36
bound 34
braid 34
hair ornament 33
black shirt 32
veins 32
cosplay 31
depth of field 31
swimsuit 31
white shirt 31
pants 30
oral invitation 30
denim 29
bondage 29
saliva 28
fishnets 28
leash 28
black bra 27
ponytail 27
tongue piercing 27
tan 27
dark-skinned female 27
head tilt 27
open clothes 27
window 27
ear piercing 27
multiple girls 26
lipstick 26
breasts apart 26
artist name 26
body writing 26
after fellatio 25
tanlines 25
couch 25
skirt 25
simple background 24
curly hair 24
eyeshadow 24
nose piercing 24
bikini 24
ejaculation 24
heart 23
bathroom 23
sleeping 23
breasts out 22
grey eyes 22
plant 22
twin braids 22
navel piercing 22
black-framed eyewear 21
sweater 21
2girls 21
watch 21
Training
I used https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb to train the LoRA. The training model was sd-v1-5-pruned-noema-fp16.safetensors. Tags are automatically shuffled and there is no activation tags (activation_tags set to 0).
Due to the size of the dataset, each image is only repeated once (i.e., no repetition). I used 5 epochs, with a batch size of 2. The UNet learning rate was set to 5e-4. The LoRA network dimension was set to 32, and the network alpha to 16.
Further notes on other tentatives
I tried to manually filter images to only include "nice" images:
high enough resolution and quality
no cum on body
"pretty" women (according to my taste)
I filtered down the dataset to 1170 images, re-trained a LoRA, but the resulting model is clearly under-performing. I even have hard times trying make cum appear on the generated images with this model.
I tried different LoRA network dimension and alpha dimension ((16, 8), (32, 16) and (64, 32)), but the best performing one seems to be the (32, 16).
Contact for collaboration
If you want to help me generate a v0.2 or even v1 of this model, please contact me at the address below by presenting what you think you can bring to the project.
For the address, everything enclosed in square braces should be replaced, spaces should be removed.
[my pseudo here in civitai] [dot] dev [at] protonmail.com