It gets better the more you use it

PIXL ships with a curated knowledge base covering color theory, dithering patterns, shading techniques, palette design, retro hardware constraints, and tiling rules. When the AI generates a tile, it pulls in the relevant techniques — not generic image advice, but specific pixel art craft.

The knowledge base covers 30+ topics with 1,300+ cross-referenced concepts. Ask for a dungeon wall, and PIXL knows about shadow placement, stone textures, mortar lines, and how WFC edges need to align.

Every time you accept or reject a generated tile, PIXL records what worked and what didn’t. Accept a tile? Its style gets added to PIXL’s reference. Reject one as “too sparse”? Future tiles will be denser.

After a few rounds, PIXL builds a style profile from your accepted work — light direction, color density, shadow depth, palette breadth. New tiles are scored against this profile before you even see them. The bad ones never reach your screen.

PIXL extracts an 8-point style fingerprint from your existing tiles: light direction, pixel density, shadow ratio, palette variety, color temperature, luminance, texture uniformity, and entropy. Every new tile is scored against this fingerprint to keep your tileset visually consistent.

Show PIXL three of your best tiles and it understands your aesthetic. New generations match your look — not a generic “pixel art” style, but your style.

Every generated tile gets an automatic quality score: readability (can you tell what it is?), appeal (does it look good?), and consistency (does it match your project?). The scores range from 1 to 5 stars and determine how often each tile appears in generated maps — better tiles show up more often.

For studios and serious projects, PIXL includes a full LoRA fine-tuning pipeline. Feed it your existing tileset and train a small model that generates tiles in your exact style. The trained model runs locally on your machine — your art never leaves your computer.

• Local training — runs on Apple Silicon via MLX, no cloud GPU needed
• Your data stays yours — training happens on-device, weights stay on disk
• Small and fast — LoRA adapters are a few MB, generation takes seconds
• Diverse training data— built-in MAP-Elites algorithm generates varied examples so the model doesn’t just memorize one layout

Got reference art you love? A sprite sheet from a classic game, a tileset from another project, or art you found online? PIXL can scan it, learn from it, and generate new assets in that exact style.

The Style Scanner is a three-step pipeline:

1. Scan — Drop your reference images. PIXL auto-detects sprite sheet boundaries, filters out junk tiles, and classifies everything by type (walls, floors, enemies, items).
2. Learn— Train a LoRA adapter on the scanned art. Takes 30–60 minutes on Apple Silicon. Your data never leaves your machine.
3. Generate— Use the trained adapter to create new tiles, walls, enemies, items — whatever you scanned. The AI produces assets that match the reference style.

The feedback loop means it gets better every round: accept the good tiles, reject the bad ones, retrain. Each version is more accurate than the last.

• Any source — sprite sheets, individual tiles, entire folders. PNG, JPG, BMP, GIF, WebP.
• Smart extraction — auto-detects background colors, tile boundaries, and quality. Bad patches are filtered before training.
• Style blending— interpolate between two reference styles. “70% retro RPG, 30% sci-fi” is one slider away.
• A/B comparison — generate the same tile with two different adapters to see which one is better.

Before generating, you can show the AI examples of your existing tiles as rendered images — not just text descriptions. It sees what your art actually looks like at the target size, then matches the proportions, shading, and detail level.