pixel-icon-lora-sdxl

Pixel Icon LoRA (SDXL)

Draws at a native 16×16 logical grid on a 512px canvas — one logical pixel is exactly 32 screen pixels. Extracting a finished icon is then sampling each block's centre, rather than downscaling and hoping.

Most "pixel art" models produce pixel-ish art: blocky-looking images with thousands of colours and no real lattice, which turn to mush when reduced to an actual small grid. This one draws a true lattice, so reduction is near-lossless.

Use

pipe.load_lora_weights("lizozom/pixel-icon-lora-sdxl")

prompt = ("yotoicon, 16x16 pixel art icon of a cat, low resolution game "
          "sprite, flat colours, hard edges, centred, black background")

Trigger word: yotoicon. 512×512, ~30 steps, guidance 7.

Try a LoRA scale of 0.6–0.8 rather than 1.0 — at full strength it sometimes over-commits and fragments the subject.

Method

Trained on 522 hand-made 16×16 pixel icons, nearest-neighbour upscaled to 512×512. At 16×16 every pixel is already one discrete colour, so pixel duplication produces a mathematically perfect lattice (measured interior spread: 0.0) that still carries the source palette and style.

Alpha thresholded at 128 onto solid black. Black is a far cleaner key colour than white for this material — near-black appears in only 6% of the source set and never above 15% of a single icon, whereas white collides with the cream highlights that are common in the art.

rank=16, lr=1e-4, 500 steps (~7.5 epochs), batch 4 × 2 accumulation, resolution 512, horizontal flip only.

No crop, no resize, no aspect bucketing. Any shift by a non-multiple of 32 misaligns the pixel grid and destroys the entire training signal — this is the one setting that matters.

Results

Flatness at logical grid 16 (lower is flatter) improved roughly 6× over the base model plus an off-the-shelf pixel-art LoRA:

Training data is not published: it is derived from a third party's artwork.