Style Learning

PIXL builds a profile of your art style and uses it to keep new tiles consistent with your existing work.

Style fingerprint

When you run pixl_learn_style, PIXL analyzes your tiles and extracts 8 properties:

Light direction

Where highlights fall — top-left vs bottom-right

Pixel density

How filled your tiles are — sparse vs dense

Shadow ratio

How much of the tile is in shadow

Palette breadth

How many distinct colors per tile

Run length

How uniform your color regions are

Hue bias

Overall color temperature — warm vs cool

Luminance

Average brightness across the tile

Entropy

How varied the color distribution is

These 8 numbers are your style fingerprint. Every new tile gets scored against it.

How scoring works

When you generate a new tile, PIXL compares its fingerprint to your session fingerprint. A score of 0.8 means "close match." A score of 0.3 means "very different style."

The threshold adapts: if your last 5 accepted tiles averaged 0.75, PIXL sets the minimum at 0.6 (80% of your average). Tiles below that get flagged before you see them.

Feedback loop

Every accept/reject decision teaches PIXL:

  • Accept a tile → its style features get averaged into your preferred profile
  • Reject as "too sparse" → future tiles aim for higher density
  • Reject as "bad edges" → future tiles get stricter edge validation
Tip
You don't need to explain your style in words. Just accept and reject a few tiles and PIXL figures out your preferences from the pattern.

Few-shot examples

The last 3 accepted tiles are stored as reference examples. When generating new tiles, PIXL includes these as rendered images in the prompt — the AI sees what your art actually looks like, not just a text description of style.

Using it

# Extract style from existing tiles
pixl style tileset.pax

# Output:
# Style fingerprint (from 12 tiles):
#   Light: top-left (0.35)
#   Density: dense (87%)
#   Shadows: moderate (15%)
#   Colors per tile: 3.2
#   Hue bias: 240° (cool/blue)