Ver NOOB
Trained based on NOOB E-PRED V1.0. Notice that some examples are generated with NOOB V-PRED 0.5.
Trigger words explaination:
Cfdoggy: doggystyle pose of futanari centaur on girl
Cfback: back view of futanari centaur
Cfonelegup: lifting one foreleg pose of futanari centaur
Cfstanding: standing pose with all four legs on the ground
Cftwolegsup: lifting two forelegs pose of futanari centaur
Cfdiphallia: standing pose of a futanari centaur with horse penis and humanoid penis
Cfundercarriage: carrying girl under the belly of futanari centaur
Cflying: lying pose of futanari centaur
Cfkneeling: kneeling pose of futanari centaur
Updated
Used PonyXL to train this one. Updated dataset to achieve better performance. It enhanced this concept on PonyXL and related models. See the examples first.
Typo issue solved. Now undercarriage is correctly spelled.
Fur color issue solved. See the examples.
Please be sure to use trigger words and hires.fix in generation. Some of triggerwords seems to overbake the images or bring bad styles when upweighting them. Reason may be I used 3d images when training. Use highres.fix to prevent this from happening.
Adding concepts: doggy, diphallia, undercarriage
About V2.0
In this version, finally I have different poses with weighted cation. Also with multires technque.
Updated keywords will prevent the models from misreading the tokens like kneeling, standing, etc.
Currently now have five keywords for different poses:
CFFB: back to the viewer;
CFKN: kneeling;
CFLU: two legs up;
CFLY: lying
CFST: standing or with single leg up.
The bad thing is, still, prompt is not sensitive, and many times the penis will mistaken for hooves or the centaur having multiple legs. You can use some horse penis lora to improve it. But it still needs dozens of trials.
Recommend using highres fix with good seeds to generate good images.
WORKS BADLY IN 512X512
Multires technique parameter:
multires_noise_iterations="6"
multires_noise_discount=0.3
Introduction
This LyCORIS is used to generate futanari centaur.
How to use it
I recommand using it in weight 0.7-0.8. It works well with other loras. But sometimes the cock will be mistaken for hooves.
Training details
Traingset is of about 220 images. They are mirrored before training. Half of them are chosen as regularization images. Total steps are about 15000.
I divided the training set into 5 subsets and tag the gesture. It is to control the results. But it didn't work, sad. If you want to give it a try, the trigger tags is the folder name below.
Regularization = true
resolution=512
batch_size=2
epoch=10
network_dim=32
network_alpha=32
clip_skip=2
Using AdamW8bits as optimizer:
lr="1e-4"
unet_lr="1e-4"
text_encoder_lr="1e-5"
Locon parameters:
conv_dim=4
conv_alpha=4