Originally inspired by Trials in Tainted Space - Gryvains - a Race of voluptuous half Dragons with plenty of curves and stunning colors.

you like Dragons?

you want Women??

plenty of curves???

Scales?????

then this Lora is for you!!!

Unlike the Gryvains from TiTs these ones tend to be more half Dragon ish with only one pair of Wings, elfin Ears and colors for hair, horns, lips and eyes linked to their scales. (saucy bits included)

The Lora loves tanned skin and if it goes too far into the dragon side simply add tanned skin or any other skin-color to the prompt.

It is also versatile enough to change the color of the scales to any color you like, even rainbow!

When no color is given the model falls back to the training data and will mix colors from red, green, blue, black and purple into stunning combos, often making transitions from one to the other.

The Lora will produce a semi-realistic/3d style of Images. If you want your images in a different Style simply pack in a Style Lora and lower the Weight from 1 to 0.8 and it will balance out.

Xavier is an amazing base model if you use it. Recommended cfg 3.5-5 with 18 to 22 steps on dpm++2m karras.

-- known issues --

• the model loves wings and sometimes the model will generate spears or rods sticking out of the character if wings are not in the prompt or in the negative.

• generating anything else than female characters is pretty hard since the training data is pure female but not impossible on lower weights.

• eyes on full body shots and some closeups are a bit messy

• in rare cases the skin can be off color and blend with the scale color when no skin color like tanned skin is given

-- source image for the dataset --

all source images were generated using perchance furry ai and natural language prompts, then ran through comfyui locally with xavier base model for upscale, fix and 3d + semi-realistic style tags. source images are available with metadata.

-- training --

trained on regular illustrious model

epochs: 10

num repeats: 4

batch size: 4

steps: 700

resolution: 1024

bucket: on

schuffle tags: false

keep tokens: 0

clipskip: 2

unet learning rate: 0.0008

text encoder learning rate: 0.00007

shedulder: cosine_with_restarts

shedulder cycles: 3

min snr gamma: 5

network dim: 32

network alpha: 16

noise offset: 0.1

the Lora made consistent images similar to source images from generation 4 onward during training with it being well balanced on generation 10.

72 source images ranging from full body to closeups from diffrent angles and poses withouth word spam preventing useless clogging of lora memory and unwanted concepts. (every tag gets baked into the lora as a trigger word!!)