AI art pipeline for game developers - live in your browser
Game art, fired until it ships.
Artkiln runs an art director in a loop: it generates, judges, and refines your asset until it clears the quality bar, then slices it into engine-ready textures. Hosted today; local app at launch.
An art department, not a generator
From first draft to shipped files: full-cycle art production for any game, 2D or 3D.
From one screen to engine-ready textures
One run in, a texture set out. The winning screen is sliced losslessly, every part gets a semantic name (Back button, Coin token, Currency panel) and ships as an individual transparent PNG - ready to drop into Unity, Unreal, Godot or any other 2D or 3D pipeline.
8.4
- “Excellent featured bundle banner with a clear mascot, candy jar, value callout, and price CTA.”
- “Strong comic rendering language: thick outlines, white sticker rims, halftone textures, saturated palette, and chunky typography.”
- “Clear HUD, back affordance, 2×3 offer grid, and bottom navigation establish an immediately usable shop flow.”
- “The six offer cards are slightly dense, making some reward-detail lines smaller than ideal at phone size.”
Every run keeps its full paper trail: prompts, rankings, duels, verdicts. This one shipped on run 3.
How it works
Anchor with your art
Tag shipped assets as Style or Layout references. They calibrate what “good” means for your game.
Generate a batch
On your ChatGPT subscription (via Codex), or with your own fal API key.
The loop judges
Triage ranks the whole batch, the best image duels the champion, a director signs off; every “ship” survives a second blind opinion. Full-batch quality signal at 2 model calls per batch.
Ship the textures
The winner ships as individual transparent PNGs, every texture named for what it is.
Made by the loop
Two games, two audiences, two rendering languages: one loop. The bar is calibrated to each project’s own shipped art, so “good” means something different in each.
8.5
8.5
8.4
7.0
ASSET SET The two-word brief “different emotions” became 10 named, transparent, engine-ready assets: Joyful Grin, Sad Pout, Angry Scowl, Surprised Gasp, Laughing, Scared, Confused, Sleepy, Proud Smirk, Love-Struck.
8.5
ship
8.2
7.4
7.1
The 8.5 was pinned as this project’s Style reference: every later run is scored against it.
ASSET SET The two-word brief “Different emotions” became 14 named expressions of the same gangster boss, background-cleaned and exported as transparent, engine-ready renders: Menacing, Smug, Mischievous, Amused, Determined, Furious, Disgusted, Suspicious, Exasperated, Guilty, Anxious, Desperate, Terrified and Hopeful.
Saturated comic candy and inky rubber-hose noir land in the same 8.3 – 8.5 band, each judged against its own shipped references.
Run it your way
Open the browser and work
Nothing to install: connect your ChatGPT subscription or an API key once and start a run. Your projects, references and runs stay in your account.
Runs on your machine
Your ChatGPT subscription or your API keys at provider prices. Free local runtimes for the expensive parts: background removal (PyMatting, BiRefNet) and 2×/4× upscale (SeedVR2). Your files never leave your disk.
The loop knows when to stop
Verdicts are ship / iterate / reject with a score calibrated against your own shipped art. Runs stop themselves on plateau instead of burning your budget.
Every run ends in files
Screen slicing, semantic naming, background cleanup, exports. 26 named textures from a single screen.
Local-first, keys-first
Already pay for ChatGPT? Generate on your subscription: no per-image bills. Otherwise your keys, provider prices, no markup. The price shows on the button before you click (~$0.08). Dry-run estimates a whole run.
FAQ
How much does a run cost?
If you have a ChatGPT subscription, runs cost nothing extra: the default Codex backend generates within your plan’s usage limits. Otherwise bring your own fal API key and pay providers directly at cost; Artkiln adds no markup. A built-in dry-run estimates the price of a run before you start.
Why not just Midjourney?
Midjourney gives you four images and a shrug. Artkiln ranks the batch, explains what’s wrong, iterates on the single most important fix, and signs off against a bar calibrated to your shipped art. You get transparent PNGs your engine opens directly, from a tool that speaks game production natively: screens, UI kits, icons, logos, banners, asset sets.
Can I run it near-free?
Close to it. Background removal and upscaling always run on your own GPU at zero marginal cost. Generation goes through fal on your own key; on a ChatGPT subscription it runs inside your plan’s usage limits, so there is no per-image bill.
Which platforms?
In the browser today - the hosted app needs nothing installed. The local app at launch: macOS and Linux, Windows via WSL2.
Which engines and file formats does it output?
Every texture ships as an individual PNG with a real alpha channel, sliced losslessly from the winning screen and named semantically. Plain PNGs drop straight into Unity, Unreal, Godot or any other 2D or 3D pipeline, with no proprietary format or exporter plugin in between.
Which image models?
Codex on your ChatGPT subscription, or fal.ai on your own key. Switch any time in Settings.
What about Steam’s AI disclosure?
Assets made with Artkiln are AI-generated content and should be disclosed like any other. Artkiln doesn’t change your obligations, just how good the disclosed assets are.
Will it be open source?
The source will be public at launch (source-available); paid convenience builds fund development.