AI sprite generator: make 2D game sprites and sprite sheets

An AI sprite generator turns a text prompt into 2D game art: characters, poses, items. To get game-ready sprites, generate each sprite or pose as its own image on a plain background, send your approved character image as a reference every time so it stays on-model, then cut each image out to a transparent PNG and pack the frames into a sheet in your sprite or game tool. With Sume, that is POST /v1/images for the art and POST /v1/rmbg-1.0/remove for the cutout.

Sume facts come from the Image API docs, the Sume API reference and API pricing, read on 2026-09-28. The reference technique itself is covered in consistent character AI image generator.

How do I make a game sprite with AI?

Work in this order, one image per sprite:

• Design the character first. Ask for a few options with n, which returns up to 4 images per call on most models today, pick one, and download it: result URLs are signed.
• Describe a sprite, not a scene: one subject, full body, side or three-quarter view, centered on a plain flat background, no shadow, no text. Name the look, such as pixel art or flat cel shading; that is a prompt style, not a setting.
• Draw every new pose from the approved image. Host it at a public HTTPS URL, send it in input_references, and say only what changes, as in "same character, now jumping". ChatGPT Image 2.5 takes up to 16 references, and Nano Banana 2 and Nano Banana Pro up to 10.
• Cut each pose out, then scale it to your game's size in your editor, as the next sections explain.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Same character as the reference, side view, mid-jump, full body, centered on a plain flat light gray background, no shadow, no text, pixel art style",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/hero.png" } }
    ],
    "aspect_ratio": "1:1",
    "n": 2
  }'

Can AI make a whole sprite sheet in one image?

You can ask for several poses side by side in one wide image, and the wide shapes exist: ChatGPT Image 2.5 takes exact sizes up to 3:1, such as 3840 × 1280, and Nano Banana 2 lists 4:1 and 8:1. But no request field sets a frame count, a grid or the spacing, so equal frames and even gaps are something you check for, not something you can ask for. For frames you can slice cleanly, generate one pose per image and let your sprite-sheet tool pack them.

How do I get a transparent background for sprites?

Cut it out after generating: on Sume, a transparent background comes from background removal, a second step after the image is made. Send the image's URL, such as the generated file's own Sume media URL, to POST /v1/rmbg-1.0/remove, and the completed result is a PNG with an alpha channel. It costs $0.0225 per image on API pricing, plus a 5.5% agent fee by default. The call itself is in remove background API, and AI images with a transparent background covers the same two-step method for any graphic.

What size should AI sprites be?

Use the size your game needs, and expect to scale down: on ChatGPT Image 2.5 the smallest exact square is 816 × 816, because a custom size needs at least 655,360 pixels and edges in multiples of 16. Image Upscale 1.0 can't shrink an image, since its factor runs from 1 to 4, so scaling a sprite down to, say, 64 × 64 happens in your image editor or game engine.

Pixel art is a style the model imitates, not a pixel grid. The API has no grid, palette or frame-size setting, so turning the image into a true low-resolution sprite with a fixed palette is work for a pixel editor.

What are the limits?

• The API makes still images. For animation, every frame is its own generation, and nothing guarantees the frames line up, so compare them before you use them.
• There is no seed: sending one returns 400 unsupported_parameter. References and a repeated description are the tools for keeping a character on-model.
• The image API returns flat 2D images, and Sume has no 3D model output.
• Result URLs from /v1/images are signed, so keep your own copies.