Writing prompts for album cover art
What actually changes an AI cover: subject, composition, palette, medium and era. Plus why 'album cover' in the prompt makes it worse.
Cover art prompts fail in a specific way: they describe the music instead of the image.
The model cannot hear your track. Describe the picture.
Five things that change the output
- Subject — what is literally in frame
- Composition — centred, off-centre, close crop, wide
- Palette — two or three colours, named
- Medium — film photograph, risograph print, oil painting, 3D render
- Era — “1970s”, “early 2000s web”, “Y2K”
Together: close crop of a hand holding a cassette, off-centre, muted orange and deep green, 35mm film photograph, 1970s.
Do not write “album cover”
It pulls the model toward generic square artwork with fake text on it. You already know it is a cover — the crop and the dimensions handle that.
Describe the image and let the format do its job.
Leave room for the title
Streaming apps overlay controls on the bottom third of your artwork during playback. If your subject sits centre-low, it gets covered.
Prompt for it: subject in the upper two thirds, empty space at the bottom.
Avoid text entirely
Models still render text badly, and misspelled artist names get releases rejected by Apple Music.
Generate the image clean and add type yourself. It takes two minutes and it will be correct.
Test at thumbnail size
Look at your cover at 200×200 pixels. That is roughly how most people will actually see it.
Busy compositions turn to noise. High contrast and a single clear subject survive. This is the single most useful test and almost nobody does it.
Generate a set, not a single
Six directions from one prompt, then push the one that works. Judging a cover in isolation is much harder than picking from a row.