Keep in mind that the model images are raw generations with no inpainting or upscales, they have the potential to be REALLY good! Check the rest of the description to review the instructions on how to get centaurs.

This is my first attempt at training, so this is not perfect. It was designed to give people who like/appreciate horses the opportunity to produce something other than the generic 'horse' that Stable Diffusion only seems capable of producing.

Arabian is the smaller, more refined/petite horse that makes up the majority of the model images, while Friesian is the thicker horse, usually black, and often ridden by knights in movies.

Breed keywords: Arabian horse, Friesian horse
Colors: bay, black, gray, buckskin, pinto,
Other: standing, posing, galloping, trotting, jumping, portrait, full body
Recommended format ex: Arabian horse, bay, black mane and tail, standing, full body, posing

Works pretty well on just the base 1 strength. Best at 1024x1024 as the majority of the data was this side. Varying results at 1280x720

VERSION 2 UPDATE:

New version includes added dataset of centaurs.

For use of the centaur, I recommend using the negative prompt:
(human legs:1.3), boots, tack, halter, saddle, horse head, watermark, signature, blurry, grainy, extra legs, extra arms
It is not 100% effective yet as I still need more data to feed it, but with the negative prompt below, I can usually get 70-90% of the generations to be correct centaurs.

This version is also much better at generating Friesians now as I've completely re-worked the dataset.