How to review AI-generated images before you ship them

You generated two thousand card illustrations over a weekend. Nice. Now figure out which fourteen hundred are safe to ship, because generation stopped being the bottleneck years ago and looking at the output never did. Here’s a practical process: what to look for, who should look, and what it should cost.

Why AI images need review at all

Two thousand assets with a 5% defect rate is a hundred bad images. Not five bad apples. A hundred screenshots waiting for someone’s review thread.

Storefronts made this concrete. Steam requires disclosure of AI-generated content at submission and shows it to players; itch.io asks creators to tag AI-generated assets. So buyers can see “made with AI” before they pay, and quality is the only thing left that separates you from the flood. One six-fingered hand in a store screenshot does more positioning work than your entire marketing plan. In the wrong direction.

The eight defects that slip past a quick look

Scrolling a folder of thumbnails catches almost none of these. They show up when someone looks at one item, full size, for a few deliberate seconds:

1. Anatomy failures. Hands first, then teeth and ears. And check the background figures: nobody ever zooms in on them, which is exactly why they’re wrong.
2. Garbled text baked into the art. Signs, labels, book spines, fake UI. Convincing at thumbnail size, gibberish at full size.
3. Style drift across a set. Item #1 and item #800 came from different moods of the same model, and it shows. Sets drift; a single image never warns you.
4. Prompt bleed. Fragments of your prompt turning up as literal objects in the scene.
5. Edge artifacts. Smearing near borders, ghost limbs at crop lines, objects that quietly dissolve where the composition ends.
6. Impossible composition. Lighting from two suns. Reflections of things that aren’t there.
7. Lookalike risk. An asset that landed a little too close to a known character, a logo, or one specific artist.
8. Context mismatches. The art contradicts the thing it illustrates: a fire spell rendered in ice blue, a dagger on the axe card.

Three ways to review, and what each one actually catches

Automated checkers cost $0.001–0.01 per image and are the right first pass. They catch what machines get technically wrong: resolution, NSFW, compression artifacts, some of the anatomy. What they can’t do is make a judgment call. A model grading another model misses style drift, context mismatches, and everything that depends on knowing what “good” means for your product.

Your own team catches everything, at art-director prices. Loaded senior time runs $30–60 an hour; at a realistic fifteen seconds per image that’s $0.15–0.25 per item in salary, paid in the most expensive currency you have: production hours. And one reviewer is one set of blind spots. There’s decades of vigilance research showing accuracy starts sliding after about 45 minutes of continuous review. Your art director is not exempt.

Human review panels split the job across several independent reviewers per item and aggregate the verdicts. Disagreement triggers extra reviews; hidden gold-standard items keep checking each reviewer’s accuracy. You get a consensus per item, not one stranger’s mood, for $0.15–0.20 per item. Meldar is built on this model, so yes, we’re talking our own book here. The math is checkable, though.

A working checklist

1. Define “ship-ready” in one sentence per asset class. “No anatomy errors, no legible garbled text, matches the house palette” beats a vibe.
2. Sample-test the rubric on 20 items. If two people disagree about what passes, the rubric is the problem, not the people.
3. Put at least three independent pairs of eyes on every item. One reviewer on a subtle defect is a coin flip. Three independent verdicts with agreement weighting is a measurement.
4. Demand per-item verdicts and fix-notes. “Batch looks fine overall” is not a review. You need to know that item #1,437 has a garbled sign in the top-left corner.
5. Track defect categories over time. If 30% of rejections are style drift, the fix is in your prompts and checkpoints, not in more review.
6. Re-review after any model or prompt change. A new checkpoint resets your quality baseline and won’t tell you it did.

What this costs in practice

For a studio shipping around 500 assets a month, in-house senior review runs roughly $75–125 in salary, plus the production time it displaces. Meldar’s Starter plan covers the same 500 items for $79 a month ($0.158 each), or $19 per 100-asset batch with no subscription, with scores and fix-notes back in 24–72 hours. Those are our prices, so run the math with your own salary numbers. The gap doesn’t really close. Automated checks stay in the pipeline either way; the only question is who makes the judgment calls.

Related reading: What is AI slop? · PickFu alternatives for testing AI-generated content