Knowledge Base
PIXL ships with a curated library of pixel art techniques. When the AI generates a tile, it doesn't guess — it pulls in the relevant craft knowledge for the specific task.
What's covered
Color theory
Palette design, color ramps, warm/cool balance, complementary colors
Dithering
Bayer patterns, checkerboard, gradient approximation with limited colors
Shading & lighting
Light direction, shadow placement, cel shading, ambient occlusion
Outlines
When to use them, thickness, selective outlines, color selection
Animation
Frame timing, squash and stretch, walk cycles, idle animations
Tiling
Seamless edges, transition tiles, autotiling rules, WFC constraints
Hardware constraints
NES, SNES, Game Boy, GBA — palette limits, sprite sizes, tile counts
Sprite design
Silhouettes, readability at small sizes, chibi proportions
How it works
When you ask PIXL to generate a "dungeon wall tile," the knowledge base:
- Searches for relevant passages about walls, dungeon aesthetics, tiling rules, and shadow placement
- Expands using a concept graph — "dungeon wall" connects to stone materials, mortar patterns, WFC edge compatibility
- Ranks results by relevance and injects the top passages into the AI's context
- Positions strategically — most relevant first, second-most last (where LLMs pay the most attention)
The knowledge base contains 30+ documents with 1,300+ cross-referenced concepts. When you search for one concept, related concepts are automatically discovered.
Concept graph
The knowledge base isn't a flat list. Concepts are linked by relationships:
- "dithering" → "Bayer matrix", "color reduction", "gradient approximation"
- "NES palette" → "4 colors per sprite", "background palette sharing"
- "walk cycle" → "frame count", "bob height", "arm swing"
A query for one concept pulls in its neighbors — so the AI understands the full context, not just isolated facts.
Expandable
Add your own documents to the knowledge base. Write markdown about your game's art style, specific technique notes, or reference material — it gets indexed alongside the built-in knowledge.