Cover art for a track: how to make it with AI

The track is ready — whether generated or recorded yourself — and the release needs a cover. It's a square the listener will see at the size of a fingernail in a track list, so it has its own rules: one strong image, a big name or no text at all, and no small details that turn into noise.

A neural network draws images like that well. It has one weak spot — letters. Ask the model to write the track title straight into the picture and it can easily drop a letter or draw a lookalike. So it's more reliable to split the work: the network draws the image, and the artist name and title go on top as real text — no typos, and you can change the font without redrawing the picture.

Track cover size

The standard that streaming services and distributors accept is a 3000 × 3000 pixel square. Smaller usually isn't accepted; larger isn't needed. You get that file from a 1500 × 1500 canvas: the cover exports at double pixel density, and the text on it stays sharp.

What to write to get a cover, not just a picture

1. Artist name and title — exactly as in the release. Or say outright that you want no text.
2. The track's genre and mood — "dark phonk", "summer indie pop", "lo-fi for studying". Genre suggests palette and typography better than any words about colour.
3. One main image — "a night highway in neon", "a girl in a hood from behind", "a broken cassette". One, not five: on a small square five objects are a blur.
4. Your own photo, if the artist should be on the cover — upload it rather than asking for a look-alike to be generated.
5. A series — if it's a single from an upcoming album, say so: the next covers can keep the same style.

Describe the track — the designer will put the cover together right here.

How to check the cover while there's still time

• Shrink it to icon size. In a streaming app the cover is seen at a couple of centimetres across. If the title doesn't read there — make it bigger, or drop it altogether.
• Name and title — one place, one typeface. Three fonts on a square read as an advert.
• No links, prices or platform logos. Social media handles, "listen on…" and streaming logos on a release cover are usually not allowed.
• Other people's imagery doesn't go on the cover. A film character, a brand logo or a recognizable photo need rights, even if a neural network redrew them. More on the risks in the chapter on copyright.

Edits without redrawing

Every cover is rendered in a browser and checked before you see it: is the text clipped, is there enough contrast, has the title got lost. After that, edits are edits to text and layout: "bigger title", "a harder font", "name at the bottom", "change the title to ‘Highway (remix)’". The generated image stays the same, and a cover you already liked doesn't drift off into a new random result.

That's especially handy for a series: one good cover becomes a template where only the title and the image change from single to single.

In the Twelver chat you can make a track and straight away put together a cover for it: the image from the network, the title as real text. Back to the guide "Create a song with AI".