I tried tweaking some training settings to see if I could make a better pixel art LoRA and oh boy did I.
This is trained on almost all of the unit sprites from Super Robot Wars W.
Recommended weight 1.0-1.5.
You must generate at 1024x1024 and then use some form of nearest-neighbor downscale 1/8 to 128x128.
The checkpoint used in my sample images is Nova Anime XL. It should in theory be able to work with other Illustrious models.
Increasing or decreasing the weight of a character's name will change how mechanized they are!
No trigger word.
Recommended: robot, mecha, chibi, standing, blue background, clenched hands, solo
You may also want to put "gundam" and "v-fin" in the negative prompt if those features are showing up when you don't want them.
Here's the training settings I think I changed from Civ's training defaults based on the metadata and my own memory:
Epochs: 10 (Probably could have gone lower)
Num Repeats: 4
Train Batch Size: 1
Resolution: 1024 (Make sure to crop and scale your sprite samples to this size, use nearest neighbor scaling, each pixel should be 8 pixels wide)
Shuffle tags checked.
Clip skip: 2
Network dim: 64
Noise offset: 0 (VERY important for pixel art I think?)