Stylization — the style contract for AI-generated game assets

Self-contained specification of the stylization layer for game pipelines. Owns everything that makes independently generated assets look like ONE game. Designed to be dropped into any game-generation skill: if your pipeline generates a single pixel of game art, it consumes this contract.

Interface in one paragraph: the pipeline derives ONE text string (the STYLE FORMULA) per game, gets it user-approved once, then inserts it byte-identical into every asset-generation prompt of every subsystem (sprites, animations, tiles, backgrounds, textures, 3D). This file defines how the formula is built, how prompts are assembled around it, which models to call with what parameters, and how style consistency is verified.

1. Why stylization exists

Every asset is produced by an independent model call — the model has no memory of previous calls. Different subsystems (and different team members' skills) generate different asset kinds. Without a shared rule the result is "three different games on one screen": photoreal floor + pixel-art hero + cartoon background.

The STYLE FORMULA is the glue: the only piece of shared context that travels into every generation.

2. STYLE FORMULA recipe

One English paragraph, 60–90 words, composed once per game right after the design brief is parsed, shown to the user for approval, stored in the pipeline's state (e.g. gdd.json.style_formula).

Concatenate five blocks, in order:

Rules:

• Block 3 is roles, not one global gamma. Environment gets its base colors; the hero gets colors that CONTRAST with that base (the hero must pop, never blend); hazards/pickups get one distinct signal hue. A single flat color list paints the whole world one gamma and kills element readability — style identity lives in blocks 1, 2 and 4, color serves readability.
• The perspective word in block 5 must match the genre: platformer / runner → side-view; top-down → top-down; puzzle / clicker → flat frontal. A top-down game with a side-view hero sprite is the most common cross-asset failure; the formula is where it is prevented. (3D pipelines: see section 8 — the perspective word is replaced, not reused.)
• If the user supplied a style reference image, describe ITS rendering style in your own precise words in block 1 — never paste image URLs into the formula.
• Do NOT put into the formula: asset-specific content (that's the per-asset description), background/keying instructions (that's the kind suffix), pixel dimensions or resolutions.

STYLE TOKEN (compressed form)

Some downstream fields are length-limited (e.g. Meshy texture_prompt ≤600 chars — the full formula usually fits; some UI fields may not). When a consumer cannot take the full formula, it takes the STYLE TOKEN: blocks 1 + 3-condensed + one accent from block 4, ≤120 chars, derived ONCE from the approved formula and then also frozen byte-identical. Example: flat vector cartoon, deep violet dungeon palette, glowing teal accents, warm amber torchlight, thick dark-plum outlines. The formula is the source of truth; the token is its compression — never maintain two competing style strings.

3. The byte-identical contract

The formula is inserted into EVERY generation prompt of EVERY subsystem without changing a single byte. No paraphrasing, no shortening, no "improving". Including single-asset regenerations during iteration. Re-deriving the formula mid-game is allowed only when the user explicitly asks to change the art style — that re-opens the approval gate and invalidates all existing assets.

Consumers and their obligations:

4. Prompt assembly for 2D assets

Every asset prompt is a concatenation of exactly four parts, in order:

<kind template> + <asset description> + <STYLE FORMULA byte-identical> + <kind suffix>

The asset description is a stable 3-4 word shorthand from the design doc (the round blue slime), never a character name.

Kind templates and suffixes

sprite (characters, objects — needs transparency) - template: game sprite of <description>, single character/object, full body visible, centered, - suffix: , on a solid uniform bright <KEY COLOR> background, no shadows cast on the background, no ground plane, nothing cropped at the edges

tile (repeating surfaces) - template: seamless tileable game texture tile of <description>, uniform pattern density, - suffix: , perfectly seamless edges that wrap horizontally and vertically, no border, no vignette, flat even lighting, no single focal object

background (full-frame) - template: game background of <description>, wide establishing view, - suffix: , no characters, no UI elements, slightly muted detail so foreground game elements stay readable, soft depth layering

texture (3D surface) - template: seamless tileable surface texture of <description>, - suffix: , perfectly seamless edges, flat even lighting, no perspective, no objects, photographed-flat appearance

ui (buttons, icons — needs transparency) - template: game UI element: <description>, single element, centered, - suffix: , on a solid uniform bright <KEY COLOR> background, crisp edges, no drop shadow outside the element

5. Key-color selection (the "green screen")

Models cannot output alpha; asking for "transparent background" produces a fake checkerboard. Transparent assets are generated on a solid key color and keyed out in post-processing.

Choosing the key color is a mandatory per-asset step, driven by the formula's palette:

1. default: bright magenta #FF00FF;
2. if the asset's own colors (its description OR the formula's role palette) are anywhere near pink/magenta/purple → bright green #00FF00;
3. if both are taken → bright blue #0000FF.

The chosen color goes into BOTH the prompt suffix and the keying script. Real shipped failure: a pink donut generated on magenta — unkeyable, the donut hole stayed pink in the game. Post-processing must also clear ENCLOSED key-colored regions (a donut hole is not connected to the image corners, so corner flood-fill alone misses it).

6. Models and volumes

• One primary generator (nano_banana_2) across asset kinds is itself a cohesion win — don't spread kinds across models without a reason.
• Pick resolution at generation time; never upscale afterwards.
• Regeneration budget: 2 attempts per asset (style drift / keying / tiling failures), then take the best attempt and compensate in code (tint, scale, 1px overlap). Style drift usually heals with a re-roll of the SAME prompt — it's sampling variance, not a prompt error; only edit the description when the content itself is wrong.
• The only legitimate full-set regeneration: the user asked to change the whole style (new formula → new approval → all assets redone). Never regenerate everything in response to point feedback.

7. Scale, proportion and coherence checks

• Every object in the design doc carries relative_scale — height in hero-units (hero = 1.0, spike trap ≈ 0.6, doorway ≈ 2.0). It drives in-game sizing and the contact sheet; it is NOT inserted into generation prompts.
• Contact sheet before asset approval: paste all sprites side by side at their relative scales on a strip of the generated tile. Check with eyes (the agent reads the image): proportions read right? hero pops against the tile or blends? A problem caught here costs one regen; caught after coding it costs a debugging session.
• Per-set style checklist: same rendering style on every asset (block 1)? outlines consistent (block 2)? palette roles respected — hero contrasts environment, signal hue only on hazards/pickups (block 3)? one perspective everywhere (block 5)?
• Final in-game sizes (2D): sprite 128px, tile 256px (power-of-two), background 1280×720. Downscale with NEAREST for pixel-art formulas (LANCZOS smears them), LANCZOS otherwise.

8. 2D vs 3D stylization — what changes

3D consumes the same formula but through different mechanics:

Full style change in 3D (user asks to restyle the whole game): new formula → re-approval gate (never skip it) → regenerate textures → re-derive engine lighting, fog and ambient from the new formula's blocks 3-4 → revisit the geometry clause. New textures on old lighting and old primitives read as "a picture slapped on the old game" — this exact failure shipped once (Roblox-style scene "restyled" to Mafia 3 by swapping textures only).

Realism ceiling (say it at the gate): without mesh generation, characters and props are primitives — a photoreal target style ("like Mafia 3") cannot be honestly met; textures will look pasted onto blocks. When the requested style implies realistic/detailed geometry, state the limit at the approval gate and offer the achievable version (gritty stylized: dark palette, warm hard light, fog, film grain) — or route through an image-to-3D mesh provider if the pipeline has one.

Everything else (byte-identical contract, palette roles, regen budget, checks) applies to 3D unchanged.