# AI book covers and album art — mood, format, room for the title > Cover art that reads at thumbnail size — portrait or square, mood over literal depiction, Image Art for painterly work, and space left for the typography. Published 2026-09-13 · eroq.ai — canonical: https://eroq.ai/blog/ai-book-covers-and-album-art A cover has one job that a good photograph does not: it has to work at the size of a postage stamp in a list of other covers, and then hold up again at full size when someone is deciding. Nothing about that is really a prompting problem. It is a composition problem, a contrast problem and a typography problem, and only two of those three happen inside [the Image studio](/studio/image). ## Pick the frame first Two of the five image formats do almost all cover work. - **3:4 — 896 × 1152.** The portrait frame. Book covers, poster art, anything upright. - **1:1 — 1024 × 1024.** The square. Album art, single sleeves, podcast covers, EP artwork. If your storefront wants a cover taller than 3:4, compose loose on purpose: keep the subject and the important shapes well clear of the top and bottom edges so a trim in your editor takes nothing you needed. That is the one case where planning to crop is sensible, and it only works if you decide it *before* rendering. The rest of the format question is covered in the [aspect ratios guide](/blog/ai-image-aspect-ratios-guide). The short version is that a native frame is composed for its shape and a cropped one never is, so never render a landscape and stand it on its end. One thing worth stating early: these are screen files. There is no upscaling step in the product, so what a cover render gives you is the native size above — fine for storefronts, ebooks, streaming art and social, and not a print file for a physical jacket. ## Mood beats depiction The instinct is to render the scene from chapter nine. It almost never works. A cover that depicts a plot point reads as an illustration of something you have not read yet; a cover that carries a mood reads as a promise. So write the prompt as an atmosphere with one anchor in it. - **One object, or one figure, or one landscape.** Not a cast. Not an event. - **A weather and a time of day.** Dusk, fog, hard noon, the blue half hour. These do more emotional work than any adjective about tone. - **A palette of two colors and one accent.** "Bone white and deep indigo with a single rust-red" is a cover. "Colorful and striking" is nothing. - **A texture.** Grain, gesso, linen, brushwork, halation. Texture is what stops a render looking like a render. Genre lives in those choices rather than in genre words. Cold blue with a small figure and a lot of empty space says literary; hard rim light, near-black ground and one saturated accent says thriller; warm grain, soft focus and a low sun says romance. You do not have to name the genre at all, and you get a better cover when you do not. ## Image Art is usually the model All three image models cost 10 credits and take references, negative prompts and CFG. For covers, the differences matter more than usual. [Image Art](/models/eroq-image-art) is the painterly, prose-faithful one — the model built to follow long descriptive prompts rather than compress them to keywords. Cover briefs are long descriptive prompts by nature, full of subordinate clauses about what the light is doing behind the thing in front, so this is the model that renders the middle of your sentence instead of averaging it away. Give it the whole paragraph and do not tidy it. Image One is the right pick when the cover should look photographed — a memoir, a true-crime title, a documentary record. Image Anime is the one for manga, light-novel and illustrated series covers where a consistent drawn style is the series identity. The full comparison is in [Image One vs Image Anime vs Image Art](/blog/image-one-vs-anime-vs-art), and the [art generator](/tools/ai-art-generator) page has more on the painterly side. ## Leave the title its room Here is the part that separates cover art from a nice picture: the typography is not yours to render, so the space for it is. eroq has no text tool, no title layer and no font control. Lettering that appears inside a render is texture that resembles type and falls apart when read. Your title, author name and any series mark go on in your own design tool — and the render has to have been built with that in mind, or you end up setting a title over somebody's face. Ask for the empty area explicitly, as part of the framing: - "generous empty sky across the upper third" - "flat unlit space along the bottom third for copy" - "subject low in the frame, quiet negative space above" - "centered composition with wide clean margins at top and bottom" And keep the model from decorating that space with fake lettering by putting it in the negative prompt — *text, lettering, title, typography, watermark, logo, signature* — which is exactly the use the [negative prompt](/glossary/negative-prompt) control was built for. ## Two prompts to start from A literary cover on Image Art, at 3:4: > A single wooden chair standing in shallow floodwater inside an abandoned ballroom, dawn light entering through a broken skylight and lying in a pale band across the water, gilt peeling from the walls into the flood, painted with loaded brushwork in bone white and deep indigo with one rust-red bloom, generous empty space above the chair, still and unresolved. A square sleeve, also on Image Art, with the title area planned into the composition: > A lone radio mast on a ridge under a bruised evening sky, seen from far below so the structure crosses the upper half of the square, low cloud snagging on the guy wires, grain and halation through the whole frame, deep teal and ash gray with a single sodium-orange light, wide quiet ground along the bottom third. Render each as a batch of four — 40 credits — and shrink the results to thumbnail size on your own screen before choosing. A cover that only works large is not a cover. ## Series, campaigns and the honest caveat If there is a second book or a second single, the look has to hold. Save the treatment once as a style Element — a written description of the palette, the finish and the light — and mention it in every cover prompt afterwards, so the series shares a grammar rather than a vague resemblance. The same trick keeps an album campaign, its singles and its social crops in one visual family. Two honest limits to keep in mind while planning. Music generation is not something eroq does — the art is the deliverable, not the record. And every render is saved privately to your library with its recipe, so you can publish a piece to [the community](/community) when you want to, and otherwise nobody sees it. ## FAQ ### Which format should a book cover be? 3:4 is the portrait frame to render at, coming out at 896 × 1152. If your storefront wants something taller, compose loose — subject and key shapes clear of the top and bottom — and trim in your own editor, because a native frame composed for its shape always beats a cropped landscape. ### Can eroq put the title and author name on the cover? No. There is no text or typography tool, and rendered lettering is unreliable texture rather than real type. Render the artwork with deliberate empty space where the title belongs, exclude lettering in the negative prompt, and set the type in your design tool. ### Which image model is best for cover art? Image Art for painterly and atmospheric covers, because it is the prose-faithful model that follows long descriptive prompts closely. Image One when the cover should look photographed, and Image Anime for illustrated and manga-style series. All three are 10 credits an image, so testing two is cheap. ### Are cover renders big enough for print? Not for a physical jacket. Renders come out at their format's native size and there is no upscaler, which makes them suited to storefronts, ebooks, streaming art and social rather than a press-ready file. Pick the frame, write the mood, leave the title its room — [open the Image studio](/studio/image) and batch four.