Type | |
Stats | 238 2,630 |
Reviews | (31) |
Published | Apr 22, 2024 |
Base Model | |
Training | Steps: 70 Epochs: 100 |
Usage Tips | Clip Skip: 2 Strength: 1 |
Trigger Words | Emolga, Patagium, Pokemon_(creature) |
Hash | AutoV2 80F0FEA3DA |
Emolga Character Lora for SDXL, because Emolga is my favorite Pokemon, lol.
Trained on ~150 images of various poses/artists/angles/ect. All images were tagged by me. Trained using OneTrainer with masked training, Prodigy, 3 repeats per image, 4 image batch size, (~70 steps per epoch) for 100 epochs. Dimension of 72, Alpha of 20. Total training time was ~20 Hours.
Use keywords Emolga, Patagium, Pokemon_(Creature) for best results. Additionally, Pokemon, Source_furry, feral, may help in some cases, but tend to introduce artifacts (Pokeball iconography, anthropomorphizing, weird anatomy, ect.) Consider using On all four, anthro, anthropomorphic, human in the negatives.
During training I referred to the yellow membrane/wing/appendage as a Patagium (which is the anatomically correct term). The AI REALLY does not like the membrane and made training this a pain, but through much trial and error, it seems to get it mostly consistent. (Except when the character has their arms folded, it will leave a hole to the background.) Images where the Pokemon was clearly flying were tagged Flying, Gliding.
No images directly including humans were in the dataset, as that caused the training to freak out, so include human characters at your own peril.
Recommended Strength is 0.8-1.0
Standard Pony Diffusion rules apply, most likely only works with Pony related models