Sign In

Dragon Portrait

366
2.3k
39
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
2,086
Reviews
Published
Mar 3, 2023
Base Model
SD 2.1 768
Training
Steps: 3,400
Epochs: 200
Hash
AutoV2
8A2C40023F

"You see, most places have waifus or lightly dressed girls. We have... dragons!"

A model trained specifically on modern portrait paintings of European dragons, based on around 80 hand-picked high quality digital art paintings from various communities. Artist names are not included in the caption to enforce a more generic art style.

The model is more or less an accident created from a failed attempt to train a more generic dragon model to improve the goofy default look of Stable Diffusion dragons, which still gave pretty bad results... except for portraits!

Since the model was trained using a slightly modified version of EveryDream2trainer, there's no real key word. The closest equivalent would be "dragon portrait", though the caption template also includes "head (shot) painting of a dragon" for better generalization.

The typical prompt roughly has a pattern like this:

<art style> portrait <art style> of a <mood> <color> dragon, <feature 1 description>, <feature 2 description>, <feature n description>, <backround description>

Example prompts:

  • digital portrait painting of an elegant cave dragon, open mouth, detailed bio-luminescent scales, emitting magic glow, in a dark gem cave

  • watercolor portrait of a cute colorful rainbow dragon, rain drops in the background

  • portrait statue of a dragon goddess from ancient egypt, ornate gold, egyptian ruins in the background

  • ancient roman mosaic portrait of an elegant dragon

  • head painting of a raging red dragon engulfed in flames, a burned village covered in ashes in the background

For better consistency, all faces are facing to the right, since this is the direction that most of the training images use. You may want to flip the results manually if you want them to face the other direction.

TIs like Midjourney style seem to work fairly well and can enhance the results, but more experimentation is needed.