Most AI album covers fail the thumbnail test: gorgeous at full size, mush at the 64-pixel row Spotify shows in a playlist. This tutorial gives you a composition scaffold, a palette discipline, and a type-overlay workflow that produces covers that still read as yours when shrunk — plus a one-hour iteration loop that beats 30 random regenerations.
The specs and tool prices below are current as of June 2026.
TL;DR
- Spotify accepts cover art from 640px to 10000px square (1:1), sRGB, 24-bit, lossless TIFF/PNG/JPG, no embedded color profile. Most distributors (DistroKid, etc.) want a 3000×3000 RGB JPG under ~10 MB. Generate at full size; never upscale a 1024px render.
- Use Midjourney V8.1 ($10 Basic / $30 Standard, as of June 2026) for the image, but add title and artist name in an editor afterward — even V8.1 still warps long text occasionally.
- The composition rule that survives 64px: one dominant shape (≈60% of frame), two high-contrast colors, a concrete subject, and intentional texture.
- Judge every variant at 150px and 64px before full size. If a cover loses its silhouette in grayscale, it loses against busy feed thumbnails too.
What this covers
A reliable workflow for AI-generated album art that survives thumbnail compression: one strong shape, two-color contrast, type that does not fight the image, and an iteration loop that gets you a final cover in under an hour instead of a long evening of regenerating.
Who this is for
Independent musicians self-releasing on DSPs, podcast hosts who keep using the default cover template, beatmakers selling on Bandcamp, and creators who need cover art for SoundCloud or YouTube uploads but cannot justify a designer for every release.
When to reach for it
Single releases, EPs, mixtapes, beat-pack covers, podcast episode art, YouTube music uploads, Bandcamp digital releases, and any time you need cover art the same day you mastered the track. Less ideal for label releases that need print-quality CMYK work — that still needs a human designer.
Pick your generator
For album art specifically, the choice in June 2026 comes down to two tools, and they fail in opposite ways:
| Tool | Plan (June 2026) | Strength for covers | Weakness |
|---|---|---|---|
| Midjourney V8.1 | Basic $10 / Standard $30 / Pro $60 / Mega $120/mo (annual ≈ $8/$24/$48/$96) | Strongest art direction, texture, film looks; native HD 2K since V8.1 (Apr 30, 2026) | Long text in-image still warps occasionally — add type in post |
| ChatGPT (GPT Image 2) | Free $0 (tight limits) / Plus $20 / Pro $100 or $200/mo | Cleanest in-image text rendering of the two | Less distinctive art direction; covers can feel generic |
The recommendation in this tutorial: generate the image in Midjourney V8.1 for the look, then set type in your image editor. If you genuinely need legible text baked in (rare for covers, common for posters), GPT Image 2 is the better text engine. Relax Mode (unlimited slow generations) starts at the Standard plan; Stealth Mode for private images needs Pro or higher.
Before you start
- Listen to your track twice and write down three adjectives — “cold, urgent, washed” — that the cover should make you feel. Generic prompts produce generic covers.
- Open Spotify, scroll to a playlist, and stare at the row at actual size. That is the test your cover has to pass.
- Pick the format up front. Spotify accepts 640px–10000px square (1:1), sRGB, 24-bit, no embedded color profile; most distributors (DistroKid and similar) want a 3000×3000 RGB JPG under roughly 10 MB. Generate at full size — do not render at 1024px and upscale, because distributors reject blurry artwork and Spotify explicitly warns against upscaled images.
- Decide whether the title and artist name will live inside the image or be added in post. Both work; mixing them mid-run wastes credits.
- Collect two reference covers from the last 12 months in your genre that you wish you had made. Name them by what works — “the negative space,” “the off-center face,” “the duotone.”
Step by step
- Start with a one-shape brief: “a single dominant shape that fills 60% of the frame.” Album covers that read at thumbnail almost always have one shape doing the heavy lifting.
- Add a two-color palette with strong contrast — “deep oxblood and bone white,” “cold cobalt and dusty cream.” Three colors is the upper limit; four turns to mud at thumbnail.
- Describe the subject concretely: not “abstract energy” but “a single figure shot from behind, shoulders to crown, against a flat washed-pink wall.” Concrete subjects compress; abstractions do not.
- Add the photographic / illustration mode: “shot on Pentax 67 with expired film,” “risograph print, two-color overlay, slight misregistration,” “oil-paint impasto, palette knife, no fine detail.” Pick a tradition; the model knows the look. In Midjourney, append
--ar 1:1so you are composing for square from the first generation, not cropping a 16:9 later. - Add the texture words last: “grain visible, slight halation, paper texture under the ink.” Without them, defaults look digital and clean — wrong mood for most music.
- Generate two grids (8 variants) from the prompt. Shrink each to 150px in your image tool and judge from the small version first.
- Pick the 1-2 strongest at thumbnail size and run targeted edits — change only the palette, only the shape, or only the texture per pass. In Midjourney, “Vary (Subtle)” holds the composition while you nudge one element; “Vary (Strong)” throws it out, so use Subtle for finalists.
A worked prompt
Assemble the brief in the order above. A complete starting prompt for a moody electronic single looks like this:
single dominant figure shot from behind, shoulders to crown,
against a flat washed-pink wall, filling 60% of the frame,
deep oxblood and bone white two-color palette, high contrast,
shot on Pentax 67 with expired film, grain visible,
slight halation, paper texture --ar 1:1 --v 8.1Note what is missing: no title text, no “vibey,” no four-color rainbow, no “abstract energy.” Every line is a concrete, compressible instruction. Swap one phrase per iteration — the palette line is the highest-leverage change once the composition reads.
First-run exercise
- Pick one track you have already released — you have the listening data and the old cover to compare against.
- Run the full prompt once and save raw outputs without tweaking.
- View at 150px and 64px. Reject anything that turns to mush. Note which composition rule each survivor follows.
- For the second pass, change only one variable — the palette is the highest-leverage swap if the composition already reads.
Quality check
- At 64 px, can you still tell what the subject is? If not, the shape is too busy or the palette has too many midtones.
- Is the contrast doing the work? Convert to grayscale; if the cover still has a clear silhouette in grayscale, it will read in any feed.
- Does the title placement avoid the subject’s focal point? Type on top of a face or hand is the most common AI-cover failure.
- Does the texture match the music? Hyper-clean digital art on a lo-fi tape album reads as off; intentional grain on a clean studio recording reads as styled, not lazy.
- Is the export valid? Confirm sRGB color space with no embedded color profile (Apple Music in particular rejects covers with the wrong profile), 1:1 square, and under your distributor’s size cap (DistroKid: 10 MB, 1000–3000px).
How to reuse this workflow
- Save the winning prompt as a template named by mood, not by release: “cold-urgent-thumbnail,” “warm-saturated-acoustic.” You will reuse the mood across releases.
- Build a small “thumbnail test” folder of 150 px exports — your visual library of what works for your audience.
- Re-test your template every 4-6 releases; model defaults shift and your texture words may no longer be needed.
- Pair with a fixed type system — one display font for title, one mono for credits — and add type in your image editor, not in the prompt. Type-in-prompt rarely reads cleanly.
- Keep a “rejects with notes” folder — covers that almost worked, each labelled with the one thing that broke them. This is faster than re-reading prompt history when you come back in three months.
- When you release a sibling track, start from the winning prompt and change exactly one element — usually palette. The cohort should feel related at thumbnail glance.
Recommended workflow
One-shape brief + two-color palette + concrete subject + medium-specific texture → 8 variants → thumbnail test at 150px → 2 finalists → targeted variable swap → add type in post. If the first 8 come back as full-frame chaos with no clear shape, the brief is wrong, not the model — rewrite the shape line before regenerating. For high-stakes singles, route the final through a quick grade in your image editor (curve adjust, slight grain bump) and export at 3000×3000 sRGB JPG under 10 MB with no embedded color profile.
Common mistakes
- No single dominant shape — busy covers turn to mush below 200px and Spotify playlist thumbnails sit around 64px
- Three or four colors at similar saturation — high midtone count is the most common thumbnail killer
- Type baked into the AI generation — letters warp, kerning fails, and you cannot reuse the artwork without retyping
- Generating at 1024px and upscaling — distributors reject blurry artwork, Spotify warns against upscaled images, and your detail is gone
- Judging at full size only — the cover will live at thumbnail 95% of the time; judge there first
- Vague mood words — “vibey,” “cool,” “aesthetic” produce the same average cover every time
- Choosing the most “beautiful” output instead of the most distinct — beauty is a low bar; recognisability at thumbnail is the bar that matters
- Skipping the grayscale test — a cover that holds in grayscale will hold in any feed background; one that does not will lose against busy thumbnails
- Treating one good variant as a finished cover without iterating — the second pass is where the cover actually becomes yours
FAQ
Which tool should I actually use for the image? For the art, Midjourney V8.1 (Basic $10/mo, or $30 Standard for unlimited Relax Mode, as of June 2026) gives the strongest direction and texture. Add title and artist type afterward in an image editor. If you need legible text rendered inside the image, ChatGPT’s GPT Image 2 (on Plus, $20/mo) renders type more cleanly than Midjourney, but its art direction tends to feel more generic.
Why does my title text come out garbled? Because you put it in the AI prompt. Even Midjourney V8.1, which improved text-in-quotes accuracy, still warps longer phrases. The reliable workflow is to generate a clean image, then add title and artist name as a real type layer in your editor.
Will distributors reject AI-generated covers? Not currently for being AI-made, but artwork is rejected for technical reasons: wrong color profile (Apple Music is strict — use sRGB with no embedded profile), non-square aspect ratio, blurry upscales, or visible logos, URLs, and prices in the image. Keep your prompt log and source files in case a platform asks for provenance, since rules are tightening.
What are the exact upload specs? Spotify accepts 640px–10000px square (1:1), sRGB, 24-bit, lossless TIFF/PNG/JPG, no embedded color profile. Most distributors standardize on 3000×3000 RGB JPG; DistroKid caps files at 10 MB and accepts 1000–3000px. When in doubt, export 3000×3000 sRGB JPG with no embedded profile and you pass everywhere.
What about cover art for vinyl? Different workflow — 12×12 inch at 300 dpi means 3600×3600 minimum, and the screen texture you want often looks muddy in print. Render print versions separately.
Square only, or also banner variants? Generate a 3000×3000 master, then crop or paint-extend it for banner, story, and poster variants. Do not regenerate from scratch — you will lose the visual identity.
How many regenerations is too many? If you are past 40 outputs without a finalist, the brief is the problem, not the model. Stop, rewrite the shape and palette lines, and restart fresh.
Should the cover match the music genre exactly? Match the mood, not the genre cliche. A folk record with a deliberately industrial cover often outperforms another acoustic-on-wood treatment in the feed.
Related
- AI brand visual direction — define the visual system before the cover
- AI image prompt basics — prompt structure fundamentals
- AI cinematic camera workflow — lens and lighting vocabulary that transfers
- AI consistent character images — keep a recurring subject across releases
- AI app background images — adjacent workflow for cover-like backgrounds
- AI ad creative tutorial — thumbnail-survival applies to ad creative too
- AI Fantasy Character Design Tutorial: From Sheet to Splash
- AI Fashion Lookbook Tutorial: Consistent Model, Different Outfits
External references:
- Spotify: Cover art requirements — the official spec your upload has to pass
- Midjourney: Comparing plans — current tiers, Relax Mode, and Stealth Mode