A letterpress print-shop world in seven sheets — key art, machine turnaround, three zones, props and materials — proven buildable by modelling the iron hand press itself: arch frame, screw and platen, rolling bed on rails, lever and rounce crank. The lab's first pure machine probe.
A Victorian conservatory world in seven sheets — proven buildable by modelling the pavilion with scan-library PBR and HDRI light. The lab's first zero-regeneration pack.
An alpine hot-spring world in seven sheets — key art, turnaround, three zones, props and materials — proven buildable by modelling the fluffy Steamfell Macaque, strand fur included.
A desert sand-sailing world in seven sheets — key art, vehicle turnaround, three zones, props and materials — proven buildable by modelling the skiff, cloth-simmed sail included.
A deep-sea salvage world in seven sheets — key art, turnaround, three zones, props and materials — proven buildable by modelling the Brass Diver straight from the turnaround.
No projects in this category. Choose another category.
AI agents create a coherent concept pack
AI agents direct image generation, compare the sheets for consistency and refine the prompts when the visual direction drifts. 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.
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 sheetmodeled · 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.
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.
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.
Is a concept pack already a playable game?
No. A concept pack establishes a visual direction and references for production. A game still needs implemented mechanics, assets, controls and testing in the selected engine.
What should I look for when reviewing a pack?
Check whether the player, environment and supporting elements share a clear palette, scale and shape language. The references should explain what to build, not merely present attractive unrelated images.
How can I communicate the art direction I want?
Describe the mood, palette, camera and important subjects, and identify useful references. Explain which visual qualities matter so the brief can guide a coherent world rather than a collection of disconnected assets.
Build a game from concepts like these
The builder draws a full pack, proves it element by element, and builds what it shows.