AI-Generated Book Covers on KDP: Fixing Resolution and Upscaling Rejections

If KDP's cover previewer flagged your AI-generated art as "low resolution" or the file preview looks soft and waxy once it's blown up to full cover size, you're running into one of the most common failure points for AI-assisted publishing: image generators are built to satisfy a screen, not a printing press. The fix isn't always "upscale it and try again" — sometimes upscaling makes the problem worse. This post walks through why the mismatch happens, how to diagnose which problem you actually have, and what a realistic upscaling workflow looks like for a full wraparound cover.

Why AI Generators Output Screen Resolution, Not Print Resolution

Midjourney, DALL-E 3, and most Stable Diffusion checkpoints generate images sized in raw pixels with no concept of physical inches or DPI attached. A default Midjourney job is roughly 1024×1024 px. DALL-E 3 through the ChatGPT interface tops out around 1024×1024 or 1792×1024 for widescreen aspect ratios. Even Midjourney's built-in upscaler typically only doubles that to around 2048×2048 px.

KDP requires 300 DPI at final print size for every cover PDF (see KDP's cover image guidelines, help page G6GTK3T3NUHKLEFX). Do the math on a native 1024 px output: 1024 ÷ 300 = 3.413 inches. That's smaller than the shortest dimension on any KDP trim size — even the smallest paperback, 5" × 8", needs a front panel of 5" × 8.25" with bleed, which is 1,537.5 × 2,475 px at 300 DPI. A raw AI export can't cover even the smallest trim's front panel, let alone a full wrap with spine and back cover included.

This is why the rejection almost always happens after the image gets placed and stretched to fill a much larger canvas in Canva, Photoshop, or a KDP template. The pixel count didn't change — you just spread the same pixels over more inches, which drops the effective DPI below 300 even though the image "looks fine" on your monitor.

Diagnosing the Real Problem: Pixel Count vs. AI Softness

Before you spend an afternoon in an upscaler, figure out which of two distinct problems you're actually dealing with. They look similar on screen but need completely different fixes.

Problem 1: True low pixel count

Open the file in Photoshop (or any editor that reports resolution), go to Image > Image Size, and uncheck resample before you read the numbers. Enter your actual print dimensions (trim width plus bleed, trim height plus bleed) and look at what DPI the software reports without resampling. If it reports something like 96 or 150 DPI at your target size, that's a genuine pixel shortage — upscaling is the correct next step.

Problem 2: AI softness that upscaling can't fix

If the reported DPI is already at or above 300 but the cover still looks fuzzy, waxy, or "melted" in fine detail — hands, small text, jewelry, fabric texture — that's not a resolution problem. It's baked-in generation softness, and it's common in AI output because diffusion models render broad shapes with more confidence than fine detail. Zoom to 100% (actual pixels) in Photoshop or Acrobat at final print size and look closely. If the mush is visible at 100% zoom, no upscaler will recover detail that was never rendered in the first place — it will only enlarge the mush or hallucinate new detail that may look plausible from a distance but strange up close.

Get Your Exact Pixel Target Before You Touch an Upscaler

The biggest mistake in fixing an AI cover rejection is upscaling to a round number ("let's just make it 4000 px wide") instead of the number KDP actually needs. Run your specs through the KDP cover calculator and enter your trim size, page count, interior ink type, and paper color. It will return your exact spine width and the full wraparound cover dimensions in both inches and pixels at 300 DPI — the number you need to hit before you upscale anything.

The height side of that number doesn't depend on your page count, so you can sanity-check it yourself right now. Full cover height at 300 DPI is always:

(trim height + 0.25" bleed) × 300

The width side is the one that trips people up, because it includes spine width, which changes with page count and paper stock. A 300-page book on white paper has a narrower spine than a 300-page book on cream stock, and both are different again on a hardcover case. Rather than guessing, let the calculator compute the spine and give you the total wrap width so you know exactly how many pixels you need across the whole panel — front cover, spine, and back cover combined, plus 0.125" bleed on each outer edge.

Upscaling Workflows That Actually Work (and Where They Break Down)

Once you know your target, here's how the common tools perform in practice:

Topaz Gigapixel AI

The most reliable general-purpose option for AI art. Its "Standard" and "Redefine" models can push a 1024 px generation to 4×–6× its original size while inferring plausible edge detail. It handles backgrounds, textures, and painterly styles well. It struggles on the same things AI generators already struggle with — hands, small faces, and embedded text — because it's built on similar generative logic and can amplify existing weirdness rather than correct it.

Photoshop Super Resolution

Built into recent Camera Raw / Photoshop versions, this gives a clean, fixed 2× upscale (double the linear pixel dimensions) with solid detail preservation for photographic content. For AI art it's a safe, conservative choice, but 2× alone often isn't enough — a 1024 px native output only reaches 2048 px, still short of most full-wrap targets.

GAN-based upscalers (Real-ESRGAN, GFPGAN)

Free and effective, particularly GFPGAN for faces. The tradeoff is a recognizable "AI sheen" — overly smooth skin, slightly plastic texture — that can compound with the same tendency already present in the source image.

Where all of these fall short: full-bleed wraps

Every upscaler above works on a single rectangular image. A full KDP wrap cover isn't one coherent scene at native AI resolution — it's a front panel, a spine, and a back panel that together can span 12–16 inches of continuous width. If your workflow is "generate one square image, then stretch and upscale it across the entire wrap," you'll get visible seams, warped proportions at the spine fold, and style drift between front and back. The more reliable approach is to generate (or outpaint) the front cover, spine, and back cover as distinct regions that share a consistent style and color palette, then assemble them at full resolution rather than stretching one small asset across the whole board.

The Judgment Call: Regenerate Native vs. Accept Upscale Artifacts

Not every project justifies chasing a perfect native-resolution render. Here's how to decide:

• Regenerate at higher native resolution when: the art is the hero of the cover — a close-up face, a detailed character illustration, a genre cover (romance, thriller, fantasy) where readers judge quality by the art itself. If your target pixel dimensions require more than roughly a 4× jump from your native output, expect visible softness or hallucinated detail in focal areas no matter which upscaler you use. Many current tools support higher native resolution than the defaults — Midjourney's "raw" mode, SDXL at 1024–2048 px native, or generating each cover panel separately at full size — and it's usually faster to regenerate correctly than to fight an upscaler for hours.
• Accept upscaling artifacts when: the AI art is a secondary or textural element — an abstract background, a soft-focus landscape behind bold title text — on a budget or non-fiction project where readers' attention goes to typography, not fine art detail. A modest 2×–3× pass through Topaz or Photoshop Super Resolution is usually indistinguishable from native-resolution art at that role, and it's the practical choice when time or budget doesn't support a full regeneration cycle.

Either way, export the final file as RGB (KDP converts internally — see paperback cover requirements, help page G201953020, and hardcover requirements, help page GDTKFJPNQCBTMRV6), flatten all layers, and keep the file under the recommended 40 MB even though KDP's hard limit is 650 MB — heavily upscaled AI art can bloat file size fast without adding usable quality.

Conclusion

AI-generated cover art fails KDP's resolution check for a predictable, fixable reason: generators output fixed pixel counts with no print context attached, and those counts are almost always smaller than what a full wraparound cover needs at 300 DPI. Diagnose whether you have a true pixel shortfall or baked-in AI softness first, get your exact target dimensions from the KDP cover calculator before you upscale anything, and match your upscaling effort — or your decision to regenerate from scratch — to how much weight the art actually carries on your cover.