AI album cover generator: square art at 3000×3000
Generate square album art, then upscale: Apple recommends at least 3000×3000. On Sume, generate 2400×2400 and upscale it 1.25× to reach 3000×3000.

An AI album cover generator makes square cover art from a text prompt or a photo: generate at 1:1, deliver at least 3000 × 3000 pixels, the minimum Apple recommends for Apple Music, and set the artist name and title yourself. On Sume, ChatGPT Image 2.5 takes square custom sizes up to 2880 × 2880, so generate 2400 × 2400, upscale it 1.25× with Image Upscale 1.0 (2400 × 1.25 = 3000), and check the size of the file you get back.
Apple's numbers and rules are quoted from its Apple Music Specification and Apple Music Style Guide; Sume facts come from the Image API docs and the Sume API reference. All were read on 2026-09-27. Other stores and distributors publish their own specs, so check the one you deliver through.
What size and rules apply to album art?
Apple's delivery spec recommends album artwork of at least 3000 by 3000 pixels, and its style guide sets content rules for the art itself:
| Rule | What Apple says |
|---|---|
| Size | Recommended minimum of 3000 by 3000 pixels |
| Quality | Not low quality, pixelated, misaligned, rotated, or with other quality issues |
| Extras | No website addresses, logos, QR codes, or references to competitors of Apple Music and the iTunes Store |
| Prices | No references to pricing |
| AI disclosure | AI-generated album artwork is disclosed with the ai_transparency tag, value Artwork |
How do I generate the album cover?
Send POST /v1/images to ChatGPT Image 2.5 (openai/gpt-image-2.5) with a square custom image_size. Custom sizes need both edges in multiples of 16 and at most 8,294,400 pixels, so the largest square is 2880 × 2880; 2400 × 2400 is the largest one that upscales to exactly 3000 with a two-decimal factor.
- Ask for no text, and describe the mood, colors, and one strong subject.
- Add
nfor up to 4 versions to choose from, the model's ceiling today; each image is billed. The price is estimated from size and quality, andusage.costis the USD amount billed. - High
quality, the default on this model, and a largenare slow configurations most likely to return202with a job to poll instead of the images. - Results are Sume-hosted, signed URLs in
data[].url, so download the cover you keep.
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": "Album cover art: a lone red rowboat on a still black lake at night, one streetlamp on the shore, grainy film look, no text",
"image_size": { "width": 2400, "height": 2400 }
}'How do I get the cover to 3000 × 3000?
Upscale it with Image Upscale 1.0: send the cover's public HTTPS URL as image_url to POST /v1/image-upscale-1.0/upscale with upscale_factor: 1.25. The factor is a number from 1 to 4, so a decimal is fine, and 2400 × 1.25 = 3000. The job runs async by default; each result artifact can report width and height, so confirm the size before you deliver it.
API pricing lists Image upscale at $0.20 per image, plus a 5.5% agent fee by default. For other starting sizes, How to upscale an image to 3000x3000 works out the factor.
curl -X POST https://api.sume.com/v1/image-upscale-1.0/upscale \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"image_url": "https://example.com/cover-2400.png",
"upscale_factor": 1.25,
"output_format": "jpg"
}'Can I make the cover from a band photo?
Yes. Send the photo in input_references as a public HTTPS URL (ChatGPT Image 2.5 takes up to 16), describe the treatment, such as the band as a halftone print on torn paper, and keep the square image_size. The model draws a new image, so faces can change; compare the result with the photo.
Nano Banana 2 and Nano Banana Pro also list 1:1 and a 4K tier. The docs don't give that tier's pixel size, so check the file before you upscale or deliver it.
Should the AI write the artist name and title?
Only if you proofread it: generated lettering can be misspelled. The safer route is to generate the art with no text, set the name and title in a design tool, and export the final square. AI image with text covers both routes.
Sources
Related posts
More in Use cases
- AI avatar for online course videos: build and update lessons
Use an AI avatar as your online course instructor: one reusable avatar, a short talking video per section, captions, and one Timeline join per lesson.
- Talking avatar for PowerPoint presentations, slide by slide
Make a talking avatar presenter for PowerPoint: one Sume clip per slide, up to 60 seconds each, in 16:9 or 4:3 to match the slide, inserted as MP4.
- AI avatar for YouTube videos: Shorts and long-form
Use an AI avatar in YouTube videos: a 9:16 talking video of up to 60 seconds for a Short, or 16:9 segments joined into one long-form video.
- AI avatar language tutor: slow, captioned lesson videos
Make AI avatar language tutor videos: speak each lesson line slowly with TTS 1.0, lip sync a tutor avatar with Fabric, then caption each clip.
Written by Sume