17/09/25 V3

I made a research on "good loras" on civitai. Most of them had ~4-6k steps, but in the end turns out it works bad with pixel art without big dataset. I mean, really big, 200+ images.

I increased size as experiment, but results differ not too much on one-to-one compare. 30 epochs were intended, but model started to overfit.

Since it is still very noisy, it is highly recommended to use pixel art loras(with limited weights) or models to stabilize pixel grid. The images were made primarily with prompts from the previous versions, sorry for being too lazy.

This model don't have that artist's charm, but still making good angular pixel lines, especially with robots, tails, wings etc. Sadly, no clear simple shapes and margin between "coloring" and outline, which was the whole purpose.

If you have suggestions on how to make training on civitai more stable, with true 4x pixel grid - please, contact me.

I have finished with this model for now. Planning to make other pixel artists loras if they do not specifically asks "no AI/NFT" in their pages.

10/09/25 V2

Used preprocessed images, more in "about" section.

There was a lot of the following tags: 1boy, 1girl, furry, mech, robot, armor, cape, coat, horns and the ones in the previous version except sprite sheets and reflections.

05/09/25 Release

LoRA trained on 51 of xlyphon's art 1024x1024. The model will be deleted immediately if requested by the author. The base model is Plant Milk Walnut (Illustrous), linked in the suggested resources.

It is trained with pixel art, xlyphon primary tags.

During training, there's often also were the following tags: robot, mecha, furry, {red, blue, orange, white} outline, simple backgroung, black/gray background, reflection on bottom, sprite sheet, armor.

This is my first attempt at any LoRA training.