How to make game assets with AI: art and sprites

Game assets are the art, sprites, models, sound and music that ship with a game. Coding agents make them in three ways: they write code that draws them, write scripts that build them in a tool such as Blender, or call an image generator.

The tools differ here. All current Claude models produce text output, according to Anthropic (model overview). Claude Code therefore draws in code, or calls an image service you connect through an MCP server.

Cursor’s Agent has image generation built in (Cursor docs), and OpenAI’s browser-game guide gives Codex an ImageGen skill for concept art, sprites, backgrounds and UI assets (Codex use case). Every tool fact on this page is as of September 2026 and links to its source.

Which route fits your game?

Draw in code for pixel-style, abstract and simple 3D looks, and generate images when you want characters or scenery with a drawn or painted look. Some games mix the two. Shatter Drift pairs custom shaders and procedural audio with AI-generated skyboxes.

Route	What it is	Examples in our catalog
Code-drawn	The agent writes canvas, shader or geometry code, and sound is synthesized in code	Hollow Crown, Metal Assault, Out Of The Box
Generated images	An image model makes sprites, backgrounds or textures that you keep as files	Exploding Hamsters, Vibe Party
Scripted 3D	Scripts build the 3D models, for example in Blender, and the game loads the exported files	Pro Skater: The Warehouse, Call of Naija

Sound follows the same split. Web Audio sources can be computed mathematically, such as an oscillator (MDN), and HALDANE-4 pairs procedural sound with one AI-generated voice-over. AI sound effects and music for games covers audio in depth, and image to 3D for games covers models generated from a picture.

What the catalog’s games do

Some draw everything in code, and some use generated images. Data note: counts come from the 404 live games in our catalog as of September 30, 2026. How each game was made comes from its entry, which draws on the creator’s own words or its repository.

Three.js is the engine of 266 of them. Phaser has 20 games, Canvas 18 and PixiJS 10.

Several entries say their art and sound come from code:

• Hollow Crown by rehan-remade: every sprite, effect and sound is produced by code, and none of it is loaded from image or audio files. Its repository marks it as made with Claude Code.
• Metal Assault by AndreaZero: the pixel-art sprites, levels, sound effects and music are all procedural. According to its repository, Claude Fable built the whole game, assets included.
• Out Of The Box by tanuu5: visuals come from shaders and generated geometry, and sound is synthesized with Web Audio.
• Pengu Smash by s1mphon1: the visuals, water effects and chiptune music are all produced in code.
• HALDANE-4 by denisbondare: ASCII graphics in a download under 5 MB, made with Cursor.

Other games use generated images. The creator of Exploding Hamsters says the art comes from AI image generation, which they felt had become good enough to produce consistent, interesting characters. Vibe Party leans on AI-generated characters, backgrounds and game assets, and its creator wanted to test how far that could go in a full party game. BeatPaws has AI-generated cats, and the skyboxes in Shatter Drift are AI-generated.

How do you get a consistent sprite set from an image model?

Fix the style before you generate, then reuse one reference for every image. Models drift. OpenAI’s guide says its model “may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations” (OpenAI).

1. Write the style down. Palette, outline, sprite size and camera angle, in one short file that you and the agent both read.
2. Save every prompt. OpenAI’s Codex guide for browser games says to save the prompts you used, in a .prompts folder, so you can generate more of the same assets later (Codex use case).
3. Reuse reference images. In Google’s Gemini API, Nano Banana 2 (Gemini 3.1 Flash Image) takes up to 4 character images to keep a character consistent, plus up to 3 style references. Nano Banana Pro takes up to 5 character images, and the limits differ by model (Google).
4. Ask for a clean background. OpenAI documents transparent output: set background: "transparent" with png or webp (OpenAI). Google’s guide doesn’t mention transparent output, so with that tool ask for a flat single-color background and cut it out.
5. Keep frames the same size. In Phaser, a spritesheet has uniform cells in rows or columns, and an atlas has frames in any size and position (Phaser docs). In Three.js, a Sprite is a plane that always faces the camera, generally with a partially transparent texture (Three.js docs), which suits the transparent images from step 4.
6. Set pixel-art rendering. Phaser’s pixelArt option sets antialias to false and roundPixels to true, which its docs call the best setting for pixel-art games (Phaser). In plain canvas, set imageSmoothingEnabled to false (MDN), or use image-rendering: pixelated in CSS (MDN).
7. Check the result in the game. Ask the agent to run the game and take a screenshot. Anthropic’s guide lists a screenshot to compare as a check the agent can run (Claude Code docs).

Keep the art from slowing the game

Set a download budget before you generate anything. The Vibe Jam 2026 rules banned loading screens and heavy downloads, and a game had to start almost instantly (jam rules).

Two catalog games show it can be done. HALDANE-4 ships as a download under 5 MB, and Shatter Drift loads at under a megabyte, according to their entries.

Formats help too. Phaser can use any image format the browser can display (Phaser docs). Three.js’s GLTFLoader reads glTF 2.0 and lists support for WebP and AVIF textures and for Draco, Meshopt and KTX2 compression (Three.js docs). OpenAI’s image API returns PNG, JPEG or WebP.

What do Steam and itch.io ask you to disclose about AI art?

Both ask, and both name art. Steam’s Content Survey lists artwork, sound, narrative and localization among the AI-generated content to disclose, and says efficiency gains from AI tools are not its focus (Steam).

itch.io’s generative AI disclosure field asks whether a project contains generative AI and which kinds: graphics, sound, text and dialog, or code. Untagged generative AI assets are no longer eligible for its browse pages (itch.io, November 20, 2024).

Copyright is a separate question. The U.S. Copyright Office said on January 29, 2025 that AI output can be protected only where a human author has determined sufficient expressive elements, and it excludes the mere provision of prompts (Copyright Office). The report covers US law, and none of this is legal advice. Selling AI-made games covers store rules and tool terms.

Templates to adapt

Replace the brackets.

Code-drawn art:

Draw every sprite in code on a [16 by 16] grid with this palette: [list of hex colors]. Write one function per character that draws it to the canvas, so I can change the palette in one place. Load no image files. Run the game and send me a screenshot.

An image model, for a sprite sheet:

A [4]-frame walk cycle for [a small robot], side view, pixel art, [32 by 32] pixels per frame, [8]-color palette, flat single-color background, the same character in every frame.

An asset budget check, for your agent:

List every file the game loads with its size in KB, then the total. Flag anything over [200] KB and tell me how to shrink it. The first load must stay under [5] MB.

Next: 3D models and sound

• Image to 3D for games and AI sound effects and music: the 3D and audio side.
• Make a game with AI: 10 prompts to copy and adapt has more templates.
• How to make a mobile game with AI covers screen sizes and touch controls.
• Vibe coding game jam shows why small downloads matter in a jam.
• How to make a game with AI and selling AI-made games cover the basics and the rules.
• How to make a 3D game with AI and the best engine for vibe coding cover the engine side.
• Browse the Three.js, Phaser and PixiJS hubs.