Game Lab · Concept art

Concept art that
survives production

Concept art is a promise about what can be built. Most AI concept images break that promise — beautiful, and unbuildable. Every pack here is stress-tested the only way that matters: an asset is actually modeled from it before the pack ships.

Try the AI Game Builder

Seven sheets, one language

A real pack speaks production’s language: key art for mood, an orthographic turnaround for modeling, three environment zones for level variety, a prop sheet at consistent scale, a material board for surfaces. One palette, one subject, no contradictions.

Pack — "Tidewatcher’s Post"7 sheets
KEY ART
TURNAROUND ×4
ZONE 1
ZONE 2
ZONE 3
PROPS + MATERIALS

Packs that agree with themselves

The coherence audit catches what makes packs useless downstream: a key-art hero whose proportions disagree with the turnaround, zones in different palettes, props at impossible scales. Sheets that disagree get regenerated with sharper prompts — and the prompt delta is the research.

Coherence audit1 regen
Subject consistencykey art vs turnaround: proportions within 10%
One palette across sheetssampled swatches match the board
Zone 2 regenerateddrifted to a cool palette — prompt gained an anchor clause
Prompt delta recorded: palette anchors now name exact swatch hexes.

The buildability probe

The decisive gate no image-only workflow has: a hero asset is actually modeled from the turnaround in Blender, rendered as a turntable, and compared back. A sheet that under-specifies — hidden back, ambiguous joints — fails the pack, not the modeler.

Probe — hero assetBuildable
turnaround sheet
modeled · turntable
Silhouette, proportions and signature details verified against the sheet.

The inventory production builds against

Every named thing across the sheets is extracted into a written element inventory — the contract the production build must satisfy, row by row, with the verifier blocking completion on missing elements. The pack isn’t inspiration; it’s specification.

ELEMENT_INVENTORY.md17 rows
01 · Tidewatcher (hero)turnaround → rigged model + sheets
05 · Signal towerzone 1 → modeled, instanced ×1
11 · Rope-and-pulley propsprop sheet → 3 variants
15 · Wet-stone materialboard swatch 3 → shader, ground family

How a piece gets made

01

Premise

An original direction with a committed palette, shape language, mood.

02

Seven sheets

Key art, turnaround, zones, props, materials — generated.

03

Coherence audit

Sheets must agree; offenders regenerate with sharper prompts.

04

Buildability probe

A hero asset modeled from the turnaround — or the pack fails.

05

Ship + port

The pack, its probe asset, and the prompt formulas that worked.

Pieces

Each card is a piece produced by the loop above — newest first.

Questions, answered

What makes a concept pack "unbuildable"?

Hidden backs, ambiguous joints, materials that read as paint rather than surfaces, proportions that disagree between sheets. Beautiful images that can’t answer a modeler’s questions — the probe exposes them before production pays the price.

What is the research output of this lab?

Prompt formulas: the deltas that fixed failed sheets — palette anchor clauses, orthographic-consistency phrasing, material-board vocabularies. Those flow back into the platform’s concept generator so every customer pack improves.

Does every game build get packs like these?

Every 3D game build gets the full seven-sheet pack with element-inventory enforcement; 2D games get a right-sized set under the same coherence rules.

Why not just generate more images until one works?

Because the failure mode isn’t ugliness — it’s incoherence and under-specification, which you only discover in production. The probe finds it in the lab instead, and each failure sharpens the prompts permanently.

Build a game from concepts like these

The builder draws a full pack, proves it element by element, and builds what it shows.

Try the AI Game Builder