This model attempts to replicate the Game Boy Advance Fire Emblem Portrait Art Style.
It's a very beautiful style. I trained this a while back but haven't uploaded it until I saw Pixel Style (GBA Fire Emblem Portrait) and remembered I haven't uploaded it! Lmao, sorry.
Anyways, I only trained this model on quarter-bodies which means it will only make portraits from the shoulders up.
The goal with this model is Pixel Perfect accuracy which is close but not quite there.
There are other quirks I need to work out not that I'm getting back into training new models.
Expect FE10 portrait art style very soon!
Description
FAQ
Comments (9)
This looks really good. How does one aim for pixel accuracy in training? Anything beyond using the same resolution and size of pixel art for data?
The method used for upscaling is very important. A lot of times you will see smudges in the dataset if upscaled incorrectly.
Another is which images you use in your training data. I have over 3k portraits available to me to use but this model only uses 150 or so images. You have to choose the highest quality and varied images.
Next is the training method. How many steps? How strongly do you train the model? Lora? Lycoris? How many layers? Do you repeat at all? All of these things matter.
Lastly, the generation method. The prompting, sampler, checkpoint, and other loras and models used in combination with this one..
It's very complex, but we'll get there eventually. No worries.
Fantastic lora. Very flexible, even for img2img. I love the idea!
Thanks! I've got newer versions coming soon. Just gotta make sure they're working well.
@MoosieMoose Will there be a 1.5 version? If that's possible even.
It looks amazing. Is a 1.5 version possible with these results?
I'll try, but 1.5 is a bit messy
Great Lora! For anyone wondering, you'll need to scale the output of a 1024 generation down to 96px to create a pixel perfect image. (or a factor of 10.666%)
What is a recommended checkpoint to use with this?