AI Game Assets: How to Get Clean Transparent PNGs

Most guides to AI game assets tell you to prompt for "a treasure chest, transparent background, game asset" and hope. That works occasionally and wastes money the rest of the time, because you pay per image and most image models cannot produce transparency at all.

The workflow that actually holds up treats the background as a problem to be solved deliberately rather than requested politely, and it batches, because six assets on one sheet cost the same as one asset on its own.

This guide covers that workflow, the edge-quality detail that separates a usable cutout from an obvious one, and an honest account of what AI still cannot make for a game.

What can AI actually make for a game?

Flat 2D work: props, icons, UI elements, item art, portraits, backgrounds, concept art and texture references. What it cannot make is a 3D model, a rigged character, an animation sequence or a tileset that tiles seamlessly — those need either different tooling or a person.

The distinction is not about quality, it is about what an image model outputs: one flat raster image with no structure. Everything that follows in this guide is downstream of that single fact.

So the useful framing is which parts of an asset pipeline are single-image problems:

• Good fits. Inventory items, shop icons, ability icons, card art, item drops, static props, key art, environment concepts, mood boards and texture starting points.
• Workable with effort. Character portraits and sprites in a painted style, backgrounds and parallax layers, UI frames and buttons.
• Poor fits. Anything needing frame-to-frame consistency, anything needing to tile, anything needing real geometry, and anything on a strict pixel grid.

How do you get a transparent background?

Two ways, and only one of them is a prompt. Either use a model with native alpha output, which is rare — GPT Image 1.5 supports it and GPT Image 2 dropped the option — or generate on a flat unused colour and key it out afterwards. Asking a model for "transparent background" without either usually returns a white box.

Both routes work; they suit different jobs.

• Native alpha, for one hero asset. A model that writes a real alpha channel gives you a cutout with no second step. GPT Image 1.5 does this, at around eight cents an image, with aspect ratios limited to square, 3:2 and 2:3.
• Flat colour plus keying, for a set. Generate on solid magenta, cut afterwards. Cheaper, works with any model, and it is the only route that lets you put several assets in one image — which is the next section and the bigger saving.

Why put several assets on one sheet?

Because you are billed per generation, not per object. Six props spaced apart on one image cost the same as one prop alone, and cutting them apart afterwards is free. The set also comes out more consistent, because the model drew all six in a single pass with one interpretation of the style.

That second benefit is the one people underrate. Style consistency across a set is usually the hard part, and generating six items in one image solves it almost by accident — the model is not being asked to remember a style across six separate calls, it is composing one picture.

The rules for a sheet that splits cleanly:

• Nothing touches. Splitting works by finding separate connected regions. Two props whose edges meet become one asset.
• Nothing reaches the border. An object touching the frame edge merges with the background region and confuses the cut.
• Wide gaps. Ask explicitly for generous empty space between items. Models default to filling a composition; you have to push against that.
• No drop shadows on the background. A soft shadow lying on the magenta is neither object nor background and will come along for the ride.
• Six to nine per sheet. Beyond that, each item gets too few pixels to be useful at final size.

Why do cut-out assets have a coloured halo?

Because the edge pixels are a blend of the object and the background. When magenta is keyed out, those semi-transparent pixels keep the magenta mixed into them, and you get a pink fringe tracing every outline. Removing it requires un-mixing the background colour from the edge, not just deleting the background.

This is the single most common reason AI-cut assets look wrong in a game, and it is easy to miss because you inspect them on a light checkerboard and ship them onto a dark scene.

A cutout that survives contact with an engine does four things rather than one:

• Samples the background from the border, rather than assuming what colour it was.
• Floods inward from the edges instead of deleting every matching pixel, so a magenta-coloured detail inside an object is not punched out.
• Ramps the alpha across the edge band rather than making a hard binary cut, which is what keeps the outline from looking jagged.
• Defringes, subtracting the background colour from partially transparent pixels so they do not carry a tint.

Here is the part that guides skip, and it is measurable. Defringing only reaches pixels the keyer classified as partly transparent. An edge pixel that was blended enough to read as solid foreground stays fully opaque and keeps whatever background colour is mixed into it — so a fringe can survive a defringe step entirely.

The six assets in this article were measured after cutting. Around 0.8% of visible pixels still carried a noticeable magenta cast, forming a roughly one-pixel rim on each silhouette, and most of those pixels were fully opaque rather than semi-transparent. On the sword — thin, so mostly edge — it was 2.5%. That is small, invisible against a light background, and clearly visible the moment the asset sits on a dark scene.

So build the check into the workflow rather than trusting the cut:

• Inspect on a dark background, not on the light checkerboard your editor defaults to. That is where the rim shows.
• Erode by a pixel if you see one. It costs a little edge detail and reliably removes the rim.
• Or pick a friendlier key colour. Magenta is maximally distinct, which also makes its residue maximally visible. A background nearer the asset's own edge tones leaves a residue nobody notices — as long as that colour still appears nowhere in the art.

Can AI make real pixel art?

No, and this is the clearest limitation in the whole workflow. Models produce pixel-art-styled images at high resolution: the blocks are different sizes, the grid wanders, edges are anti-aliased and the palette runs to hundreds of colours. It looks right in a blog post and falls apart in a sprite editor.

Actual pixel art is defined by discipline the model has no concept of: one uniform grid, a fixed small palette, deliberate placement of every pixel, and hard edges with hand-controlled dithering instead of anti-aliasing.

You can drag AI output partway there — downsample to the target sprite size with nearest-neighbour, quantise to a limited palette, then clean up by hand — but be honest about what that is. It is a reference you are redrawing, not an asset you are importing. For painted, vector or illustrated art styles the models are genuinely useful; for pixel art they produce a mood board.

How do you keep a set looking like one set?

Generate together wherever possible, and where you cannot, anchor to an approved image rather than to a prompt. A style described in words drifts between generations. The same style supplied as a reference image holds far better, which is why sheets and reference-guided editing beat re-prompting.

Style consistency across dozens of assets is the difference between a pack and a pile. Three techniques, in order of how well they work:

• One sheet, many items. Best results, because there is only one interpretation of the style. Use this for anything that can share a frame.
• Reference the first sheet. When you need a second batch, pass the approved sheet as a reference image and ask for new items in that exact style. Reference-guided models hold a look far better than a re-typed description.
• A fixed style block in the prompt. The weakest option, but useful as a baseline: write one paragraph describing the style — medium, lighting, outline weight, palette, level of detail — and paste it unchanged into every generation. Change one thing at a time and never rewrite it mid-project.

The related problem of keeping a character recognisable across many images is a different discipline with its own techniques, covered in the character consistency guide.

Which model suits which asset?

For prop and icon sheets, the cheapest model that draws clean shapes wins, because you will generate many sheets and discard most. Save the expensive models for the two cases that need them: anything with text on it, and anything that has to match an existing reference exactly.

Rates from Rangy's live pricing tables at the cheapest configured provider, checked August 2026. Providers change rates, and resolution changes the price for models billed by resolution.

What does an asset pack cost?

Around a cent per asset if you batch, and six times that if you do not. The sheet in this article cost five cents at 2K and the cut returned six trimmed PNGs for nothing, which works out at $0.008 an asset. The same six generated individually would have been thirty cents. Neither number includes the sheets you throw away.

That last sentence matters more than the arithmetic. Realistically you generate three or four sheets to get one good one, so budget for the discard rate rather than the ideal case — a usable set of forty assets lands somewhere around a dollar rather than thirty cents, and that is still cheap enough that the cost stops being a factor in the decision.

Where the money actually goes at volume is iteration, not generation. Which is the argument for paying per image rather than a monthly fee when the work is project-shaped: heavy during a content push, nothing between them, the same reasoning as in the guide to generating without a subscription.

What can AI still not do for games?

Four things, all structural rather than fixable with a better prompt: 3D models, animation frames that stay consistent, tilesets whose edges actually match, and pixel art on a real grid. An image model returns one flat picture with no structure, and none of these are one flat picture.

Stated plainly, because this guide is published by a company that sells image generation:

• 3D assets. Image models output images. Getting geometry, UVs and a rig requires an entirely different class of tool, and generated concept art is a reference for a modeller rather than a substitute for one.
• Animation. A walk cycle needs the same character in the same style across every frame, with controlled changes between them. Frame-to-frame consistency is precisely where generative models are weakest.
• Seamless tiles. Nothing in a model guarantees that the left edge of a texture matches its right edge. Tiles need to be made tileable afterwards, and often are easier to author directly.
• Pixel art. Covered above — you get the aesthetic without the grid.
• A coherent art direction. Models execute a style; they do not decide one. A pack that looks assembled rather than designed is usually missing a person, not a better model.

The pattern that works is using generation for the wide, cheap part of the pipeline — exploration, props, icons, concept work — and keeping human effort for the parts where structure and consistency are the entire requirement.

Frequently asked questions

The bottom line

Generate sets rather than singles, on a flat colour nothing in the art uses, with wide gaps and nothing touching the frame. Cut with a keyer that ramps the alpha and defringes. Expect a reference rather than a finished asset for pixel art, animation and anything 3D.

Almost everything that separates a usable AI asset pack from an unusable one happens outside the prompt. The prompt decides what the props look like; the sheet layout decides what they cost, and the cutout decides whether they survive being dropped into a scene.

Get those two right and the model choice matters far less than it seems to.