Type | |
Stats | 3,392 1,694 |
Reviews | (489) |
Published | Mar 9, 2023 |
Base Model | |
Training | Steps: 14,400 Epochs: 10 |
Trigger Words | pokemon normal fire water grass flying fighting poison electric ground +10 more |
Hash | AutoV2 1B49190E07 |
A LORA trained on official artworks of all the 1008 Pokémon, along with their types and BLIP generated captions.
This is my first fully trained LORA model and I have no idea what I'm doing.
Any feedback or tips on how to improve this would be really helpful :)
Fell free to experiment with it and share!
Recommended Configs:
My default prompt is:
pokemon, [type1] and [type2], [characteristics], masterpiece, high quality, best quality, high-definition, ultra-detailed, black background,<lora:pokemon:0.5>
Negative prompt:
nsfw, human, 1boy, 1girl, (worst quality, low quality:1.4), ( jpeg artifacts:1.4), (depth of field, bokeh, blurry, film grain, chromatic aberration, lens flare:1.0), greyscale, monochrome, dusty sunbeams, trembling, motion lines, motion blur, emphasis lines, text, title, logo, signature
The weights could be between 0.5-0.7 for animal like and 0.2-0.5 for humanoid like Pokémon.
Clip Skip: 2 and ENSD: 31337.
768x768 works better than 512x512.
For sampler Euler a works fine, but I prefer DPM++ SDE Karras.
Use "black background" tag to keep the Pokémon isolated, more complex backgrounds are better with negative embeddings.
You can use either one or two types.
Describe your Pokémon along with vibe characteristics (Ex: edgy, cute, strong, etc).
I tested it on both AbyssOrangeMix3, and ReV Animated and get good results, realistic models such as Realistic Vision seems to not work at all, but further tests are welcomed.
Abyss has more official artwork style and ReV is more stylized.
You can get deformed creatures some times, so a negative embedding should help.
Examples of negative embeddings:
Here is a image with every type applied in the same prompt: https://imgur.com/a/pkqj5ST
The base model used for the training is the AbyssOrangeMix3, therefore the licenses of this LORA are the same.