This LoRA was trained using a single pixel art image posted by gemumeku. The original image was separated into 47 individual characters, then upscaled 10× from 240×320 using nearest-neighbor interpolation to create the training images.

For training, I also experimented with a modified version of beautifulrem's Anima-Standalone-Trainer with an implementation of the CAME optimizer.

The resulting LoRA appears to have learned the concept of pixel-perfect rendering and the proportions of the characters to some extent. When generating at higher resolutions, it tends to preserve the original character proportions while adding more detail.

The pixel-perfect reproduction is not completely accurate, however. The generated pixels often extend beyond the intended grid, and diagonal lines may occasionally appear.

These artifacts can be removed using the pixelOE node: pixelOE node example