Game art that knows a coin from a character

Ask an image model for a “character sheet” of a pit and it will hand you a character — that is what the word is for. One QA run of Game Studio did exactly that: a western called Saddlebound came back with a person whose torso was a pit, a coin standing upright in little boots, and a saddle rendered already strapped to a horse nobody asked for. The model was not malfunctioning. The word character is an instruction to a model's people-drawing prior, and the studio was saying it about things. The fix was not a better model or a sterner prompt at the end. It was teaching the whole pipeline the difference between somebody and something, and never letting the two share a sentence again.

Somebody or something is a backend fact

The cast list in your game's plan is now grouped into living members and objects, and the boundary is not guessed from names or drawn by the page. It comes from the backend's own vocabulary of which roles carry a face — the player, an enemy, an NPC are somebody; a pickup, an obstacle, a prop are something — and it ships on the same payload the art pipeline reads, so the page can never grow a second copy of the rule that quietly drifts from the one the prompts use.

The split also knows when to stay out of the way. It only surfaces when it actually separates two groups; a roster that is all characters, or all objects, stays one flat list with no headings, because a heading over everything is a heading over nothing.

The word “character” is an instruction

Every prompt on the object side of the boundary carries the same guard, in plain words: draw the object itself, exactly as described, with no body, limbs, face or eyes added; not carried, worn, ridden or mounted by anyone; not attached to any other creature or thing; nothing and nobody else in frame.

That sounds like overkill until you watch what happens without it. “Game sprite of Pit” alone reads as a character brief, and the sprite is the art the player actually meets in play — which is where a pit with a face is at its most wrong. The guard is restated everywhere the object is drawn: on its reference sheet, on every state pose, and on the sprite itself, because a rule that only some prompts carry is a rule the other prompts break.

Characters act; objects show states

There is no fixed menu of actions in Game Studio. What a character does is part of your game's scenario, so the concept planner extracts each one's actions from the brief and the design as short verb phrases — gallops over gaps, throws acorns — and the actions stage renders one pose per phrase. That much was already true when a one-line brief became a playable game.

Objects broke that frame. A coin does not do anything, yet it still has to animate. So an object's actions are reframed as the states it shows across play: a resting state first, then one state per way play changes it — spins and glints, waiting, then bursts into sparks as it is collected. The planner is pointed at how such objects usually animate in games like yours, which is why a pit yawns and swallows rather than performing gymnastics. Each state becomes a drawn, animated pose later, exactly as a character's verbs do.

The one redescription that cannot drift

Every pose is redrawn from the cast member's approved reference image, and every redraw prompt opens with an identity clamp: same species, body shape, design, clothing, colours, proportions and scale. The clamp alone turned out not to be enough. Edit models under-weight their reference more than anyone would like, and QA watched the same approved hero come back as two visibly different people two poses apart.

The repair reverses an old rule. The studio used to avoid redescribing a character in pose prompts, because a fresh description is an invitation to a fresh design. Now each pose restates the member's own approved description after the clamp — the exact words the reference sheet was drawn from. That is the one redescription that cannot ask for a different design, and it anchors the drifting model to the identity it was already supposed to hold.

Objects get the same clamp said of a thing: identical design, shape, materials, colours, proportions and scale, and no body grown along the way. A saddle in its waiting to be claimed state must not render already on a horse — that is the state answered wrongly, and the prompt says so.

A pose is an animation frame, not an illustration

Every pose is asked for as one frame of that action's animation: same camera, same scale, only the pose — or the state — changes. The reason is mechanical. A sprite here is two frames, and the runtimes loop the movement pose against the resting design, so a pose that cannot sit in a loop with its sibling is a pose that judders in play. An illustration wants drama; a frame wants continuity; the prompts ask for the frame.

The hero stays out of the scenery

Environment art had the mirror-image bug. Backdrop prompts used to carry the game's pitch, for cohesion — and a pitch is a sentence about the hero doing things. So backdrops arrived with the rider and horse painted in at foreground scale: the exact figures the uninhabited-scenery guard forbids, because a negative guard loses to a positive description every time. Environment prompts are now told only which game they serve, never what the hero is doing in it.

The stakes are practical rather than aesthetic. Sprites are drawn over the backdrop at their own scale, so anything the painting puts in the foreground fights them — a rider the size of a house is a character the player can never meet, and testers file it as a broken game object.

A model per stage, not per app

Each stage's panel now carries its own model picker, offered only the models whose kind can actually serve that stage — a text model can plan the concept, an image model cannot, and an image model asked to anyway is a failed call and a bill. The shared catalogue is fetched once, your choice is stored on the project keyed by stage, and a stage that runs no model shows no picker at all, because automatic routing needs no control. It composes with your account-wide spend tier: the picker decides where a stage starts, the tier decides how far a failure may climb.

A worked example: Saddlebound, twice

The first Saddlebound run asked one vocabulary to draw everything. Cast: a rider, a horse, a snake, a coin, a pit, a saddle. Result: the rider was fine, the snake was fine — and the coin wore boots, the pit grew a torso, and the saddle came pre-mounted, because every one of them had been asked for as a character.

The same pitch through the split pipeline: the rider and snake land in Characters with verbs extracted from the scenario — gallops over gaps, coils to strike. The coin, pit and saddle land in Objects with states — spins and glints, waiting / bursts into sparks; yawns open / swallows the unwary; waiting to be claimed / cinched and ready. Each object's sheet, states and sprite carry the thing-not-somebody guard; each character's poses carry the clamp plus its own approved description. The backdrop is told it serves a western, and nothing else. Nobody grows boots.

Why not just prompt harder?

The alternative is the one every standalone generator offers: a single prompt box, where you remember to type “no face, no body” yourself, paste your character's description into every request by hand, and keep the hero out of your background prompts through personal discipline. It works right up until the fortieth image, which is precisely when it stops: one forgotten paste is a stranger in your sprite sheet, and one pitch-flavoured backdrop is a second hero painted into the scenery.

The difference here is that the rules are data, not discipline. The boundary between somebody and something is a table the pipeline reads, the guards are constants every composer spends, and the description that anchors identity is the approved one on file — the same provenance posture as reproducible workflows, applied to prompts. The same staged shape runs the film pipeline next door; what Game Studio adds is a vocabulary for the cast members that were never people.

Frequently asked questions

How does Game Studio decide what counts as an object?

By role, not by guesswork. The backend defines which roles are somebody — the player, enemies, NPCs — and everything else on the roster is something. That vocabulary ships to the page on the same payload the prompt composers read, so the grouping you see and the prompts a generator spends can never disagree. If your roster turns out to be all one kind, the split stays invisible and you get one flat list.

Why do other tools give my pit a face?

Because the request said “character”. Image models carry a strong people-drawing prior, and the word character aims it; asked for a character sheet of a pit, the model helpfully supplies the missing person. Game Studio's object prompts name the thing, describe it exactly, and explicitly forbid a body, limbs, face or eyes — and repeat that guard on the sheet, every state pose, and the sprite, so no single prompt is the only thing holding the line.

What exactly are object “states”?

The ways play changes the thing, drawn as poses. Every object gets a resting state first, then one state per change: a coin waits, then bursts as it is collected; a pit yawns, then swallows. Each state renders as one animation frame from the object's approved reference, under a clamp that keeps the design identical — so the runtime can loop the state against the resting frame without the object turning into something else mid-loop.

Do I have to pick a model for every stage?

No. Every stage routes automatically if you leave it alone. The picker exists for when you care: it sits at the top of each stage's panel, offers only models whose kind can serve that stage, and remembers your choice on the project, per stage. A stage that runs no model — or whose kind nothing in the catalogue can serve — simply shows no picker. Your account's spend tier still applies underneath whatever you pick.

Will my existing game re-render differently now?

Nothing regenerates by itself — art you have approved stays approved. The split changes what new renders are asked for: object rows get object prompts, action poses carry the approved description as a second identity anchor, and backdrops stop hearing about the hero. If an old asset has the classic symptoms — a face on a pickup, a second hero in the scenery — re-running that one stage is the fix, and only that render is charged.

Start from a sentence in Game Studio, or walk the whole pipeline first in the Game Studio tutorial — the cast panel will show you who is somebody and what is something before a single credit is spent.