Lilly Tech Systems
Procedural Generation Best Practices
Ship quality procedural content by following proven patterns for seeding, validation, performance, and player experience design.
Seeding and Reproducibility
- Always use seeds: Every procedural system should accept a seed parameter. This enables reproduction of bugs, sharing worlds between players, and deterministic testing.
- Hierarchical seeding: Derive sub-seeds from a master seed for different systems (terrain, dungeons, items). This way, changing the item seed does not affect terrain generation.
- Display the seed: Show the seed in the UI so players can share interesting worlds (like Minecraft).
- Version your generators: Changing the algorithm breaks old seeds. Version your generators and keep old versions available when possible.
Validation and Testing
Automated Validation
Run your generator thousands of times with different seeds. Check every output for basic requirements: no inaccessible rooms, no overlapping geometry, all paths reachable.
Constraint Checking
After generation, verify constraints: minimum room sizes, maximum corridor lengths, required features present, difficulty curves maintained.
Visual Inspection Tools
Build tools to rapidly preview generated content. A gallery mode that generates and displays 50 levels at once helps spot patterns and problems quickly.
Regression Testing
Store reference outputs for known seeds. When you change the generator, compare against these references to detect unintended changes.
Performance Tips
| Technique | Description |
|---|---|
| Generate async | Run generation on background threads to avoid blocking the main thread |
| Lazy generation | Only generate what the player can see or will soon reach |
| Caching | Cache generated chunks/rooms. Do not regenerate content the player has already visited |
| LOD generation | Generate coarse versions first, add detail as the player approaches |
| Budget limits | Set time budgets per frame. Spread heavy generation across multiple frames |
Player Experience Design
- Guarantee quality minimums: Every generated level should be completable and fun. Reject or regenerate outputs that fail quality checks.
- Mix handcrafted and procedural: The Spelunky approach — hand-designed room templates assembled procedurally — often produces the best results.
- Control difficulty curves: Do not let randomness create impossible difficulty spikes. Layer difficulty scaling on top of procedural generation.
- Avoid sameness: Pure random generation can feel repetitive. Add landmark features, unique events, and authored content to break up the generated material.
- Player agency: Let players influence generation (choose biome, difficulty, seed) for more engagement.
Combining Techniques
- Multi-scale: Use different algorithms at different scales (noise for terrain, BSP for dungeons, templates for rooms, WFC for decoration).
- Pipeline approach: Chain generators: first generate layout, then populate with enemies, then place loot, then add decorations.
- Post-processing: Apply cleanup passes after generation to fix edge cases, smooth transitions, and add polish.
Validate after generation. Flood-fill from the start to verify all required areas are reachable. For lock-and-key puzzles, verify the dependency graph is solvable. Run pathfinding to confirm all objectives are accessible. If validation fails, regenerate with a different seed or apply fixes.
Start with random room placement connected by L-shaped corridors. It produces good results, is easy to implement, and teaches the fundamentals. Next try BSP for more structured layouts, then cellular automata for caves. WFC is more advanced but very rewarding once you understand constraint propagation.