Reference

Glossary

Plain-language definitions for terms the map generator surfaces.

B

Bhattacharyya coefficient: A score for how similar two colour palettes are. 1.0 = identical distributions, 0.0 = nothing in common. Used as a first seam check: if two tiles’ edge strips score ≥ 0.40, their colours blend invisibly at the join.
Biome: A map region type (forest, dungeon, desert, stone, water, etc.) that determines visual style, texture character, and validation parameters for tile generation.

ControlNet: Neural-network conditioning that provides structural guidance — edges, depth maps, or reference images — to the diffusion model during tile generation. The painter uses ControlNet Tile / Union to lock structure while the prompt drives content.

Denoise strength: Value 0.0–1.0 controlling how much the diffusion model can change each pixel. Low values preserve existing content; high values allow full regeneration. Painted seams turn this into a per-pixel schedule.

Edge strip: Narrow band of pixels (typically 64–128 px) extracted from a tile’s edge for comparison with neighbours during validation.
Exterior anchor: A 256-px strip of an already-rendered neighbour tile that a new tile is generated against. Locks the shared edge without locking the new tile’s interior.

Gaussian feather: Soft alpha falloff applied at tile overlap regions so neighbouring tiles fade into one another rather than meeting at a hard line. The painter uses a Gaussian-shaped weight kernel sized to the overlap width.

IP-Adapter: Style-conditioning mechanism that locks generated tile content to match a reference image’s aesthetic, ensuring palette and texture consistency across the grid. ipa = 0.6 is the empirical ceiling — higher values cause CLIP to replicate content rather than transfer style.

Laplacian pyramid: Multi-scale image decomposition that separates an image into frequency bands at different scales. Used for multi-resolution blending of tile overlaps.
Latent normalisation: Per-tile latents are normalised against the anchor tile’s statistics before decoding, so same-biome tiles share a tone baseline and don’t drift apart visually.

Overlap: Shared pixel region between adjacent tiles where blending occurs. Overlap width determines the transition zone for feathering and harmonisation.

Painted seam: A user-authored brushstroke (river, road, biome transition) drawn on the canvas. The constraint compiler turns it into per-tile masks that drive differential diffusion at generation time.
Palette drift: Gradual colour shift that accumulates as tiles generate further from the seed tile. Detected by comparing histogram statistics against the seed’s reference palette.
Palette match: Post-process step that runs a CPU histogram match between a finished tile and its anchor neighbour. Sub-100 ms, free win on the seam quality gate.
pHash: Perceptual hash using DCT (discrete cosine transform) for similarity-invariant image fingerprinting. Tiles with similar visual content produce similar hashes — used to flag accidental tile repetition.
Poisson blending: Gradient-domain compositing technique that preserves interior texture while seamlessly matching boundary conditions between regions.

Reinhard transfer: Closed-form colour normalisation that matches the mean and standard deviation per LAB channel, aligning the colour distribution of one image to another.

Seam: Visible boundary where two adjacent tiles meet, caused by mismatched colours, textures, lighting, or structural elements across the tile border. Seams are the central failure mode the painter is engineered to avoid.
Seed tile: The first generated tile that establishes the style, colour palette, and texture baseline for the entire grid. All subsequent tiles reference it.

Tile: A single image segment (typically 512–1024 px) in a grid-based map. The fundamental unit of generation, validation, and assembly.
Tier 0 / 1 / 2: Graduated validation levels. Tier 0: pixel-level checks only (free, < 10 ms). Tier 1: VLM seam coherence on failure. Tier 2: full VLM narrative and constraint compliance.

VLM: Vision-Language Model used for semantic evaluation of tile content, seam coherence, and constraint compliance in Tier 1 and Tier 2 validation.

Wave: A set of tiles at the same BFS distance from the seed tile. Tiles within a wave have all dependencies met and can be generated in parallel.