This is a failed experiment, don't use it.

I tried to put several concepts into one LoRA, but as you can see, the results are not good. My conclusion is that finetuning or LoRA training with dim 128 or more is needed to train several concepts that are far apart from each other.

_______________

The concepts included in this LoRA are:

robot, cyborg

(The "Robot" tag generates a subject with a more angular appearance, mechanical features such as a vent, LEDs, etc., while the "Cyborg" tag generates a subject with only robotic joints.)

species

-dragon (including western dragon)

---anthro, feral, taur

-eastern dragon

---anthro, feral, legless

-dinosaur

---anthro, feral

-crocodilian

---anthro, feral

-shark

---anthro

-and other species

nsfw concepts

-slit play

---slit fingering, spread slit

-penis tip (tagged as "pentip")

-erection, half-erect, flaccid

So the available tags are "feral, legless, dragon, western dragon, eastern dragon, dinosaur, crocodilian, shark, slit play, slit fingering, spread slit, pentip, half-erect, 3/4/5 fingers, 2/3/4/5 toes".

_______________

known issues:

Overfitted, this LoRA basically generates anime style images.

Because the dataset is so large (about 2500 images), the tagging is not perfect. You may encounter tag mix-ups, e.g. if you use "office" you may get a simple background, if you use "mouth closed" you get an open mouth, or if you use "hands above head" you get a hand on its own leg.