This is a LoRA that attempts to improve the somewhat cartoony style of dragons generated with PonyDiffusion XL and gives it a more professional high quality digital artwork style. It was trained on a selection of high-quality, high-resolution non-anthropomorphic (or what some folks call "feral") western dragons from various artists. The dataset is mostly that of my older Dragon Portrait model for SD2 with lots of additions and improvements.

PD XL was picked as a base model, since it already has a fairly sophisticated understanding of dragon anatomy and pretty good prompt understanding.

Some general tips:

• Like the base model, the LoRA was trained on a combination of natural language and tags, so it's recommended to use both together. There aren't really any keywords, the closest equivalent is probably just "dragon".

• Use "male" for more masculine or "female" for more feminine looking dragons.

• To control the perspective, use tags like "front/back/side view".

• To generate biped dragons, try to use "wyvern" instead of "dragon". The dataset currently doesn't have a lot of wyverns in it, so the results may not be as good.

• Start the prompt with "portrait of a" together with a typical portrait aspect ratio to generate portraits. Try to use "fullbody" as positive or "portrait" as negative if portraits are unintentionally generated.

• Use just "wings" for a neutral pose and "spread wings" for a more epic pose. Not including "wings" usually produces wingless dragons. "folded wings" can be used to show the wings resting and folded together, but it is a bit quirky.

• To further amplify the LoRA, it's recommended to use tags like "feral", "reptile" and "detailed scales" in the positives and "anthro" in the negatives. Though it can be used for anthropomorphic characters to a certain degree as well, of course.

• While the LoRA was not explicitly trained on NSFW art, the base model together with other LoRAs should satisfy any needs in that direction.😉