QA checklist

Background Removal Quality Checklist

A production QA checklist you can run on every batch of background-removed product photos before they reach a listing. Twelve checks, each of them worth about three minutes to catch and a reshoot to miss.

By Seb RodriguezPublished February 28, 202610 min read

TL;DR

AI background removal is 95% done out of the box. The last 5% is what shoppers notice. Run this twelve-point checklist on a random sample of every batch before you publish — your return rate and your listing suppression rate both depend on it.

Why QA matters more in 2026 than it did in 2022

Early AI background removers had obvious failure modes. You could see the bad cutouts. Edges were jagged, hair vanished, glass objects turned opaque. The tool was the bottleneck, and the bottleneck was visible.

Modern removers (including CatalogCut's model) fail less visibly. The errors are subtle: a one-pixel halo around the product edge, a slightly inaccurate hair strand, a missed reflection on a watch face. Shoppers see those errors even if they don't consciously flag them — and they translate to lower trust, lower conversion, and higher return rates.

That makes QA the most underinvested part of most sellers' photo workflows. This checklist is the minimum viable QA pass — twelve checks, running on a random sample from every batch, with a clear accept/reject criterion for each.

How to sample a batch for QA

You do not need to QA every photo in a batch. Sample 10%, minimum 5 photos, maximum 20. Pick the sample randomly — not the first five. First-photo bias is real, and the hardest failures usually hide in the middle of a batch where you've stopped looking.

If the sample passes all twelve checks, ship the batch. If any sample image fails, inspect the whole batch for the same error mode. Most failures are systematic — one preset setting is off, or one lighting condition in the original shoot is tripping the remover.

For categories with known remover weaknesses (hair, fur, glass, transparent plastic, sequins, metallic highlights), bump the sample rate to 25% and increase acceptance strictness. CatalogCut's preset system has per-category tolerance presets for these edge cases.

The twelve-point QA checklist

Run through all twelve on every sampled image. Each takes roughly 15–30 seconds of visual inspection. If you have a two-monitor setup, put the original on one screen and the processed output on the other for direct comparison.

The twelve-point QA checklist

Edge cleanliness at 100% zoom
Zoom to 100% on the product edge. No halo, no fringe, no jagged staircase pixels. A 1px halo on a 2000px image is easy to miss at screen size but obvious at retina zoom.
Hair, fur, and fine details preserved
Fine details are where removers fail hardest. Zoom to hair strands, fur wisps, feathers, mesh, or lace. If you see blocky transitions or missing strands, the remover settings need softer edge feathering.
Transparent and translucent elements handled
Glass, bottles, acrylic, water, smoke, and semi-transparent packaging require alpha-aware removal. Check that you can see through the transparent part appropriately — not a hard white cutout.
Shadow is on the correct side
If the original photo has light from the upper-left, the shadow should fall to the lower-right. Generated shadows that point the wrong way are the single most recognizable sign of AI processing.
Shadow intensity matches the product category
Jewelry and small accessories: subtle, soft shadow. Apparel: natural soft shadow. Electronics and appliances: clean drop shadow. Shadow that feels heavier or lighter than the product category is a trust signal of poor editing.
No color shift on the product
Compare the product's dominant color to the original. Removers sometimes shift color by 2–4 points on the hue wheel. For brand color accuracy (especially in apparel and cosmetics), use a color-check card in the original shoot.
No reflection artifacts
Reflective products (watches, glasses, polished metal) keep reflections of the original environment. Check that those reflections still make sense against the new background — a bathroom-sink reflection on a luxury watch against pure white is a red flag.
Product scale is consistent with other listings
If this is a variant or part of a series, verify that the product sits in the same frame coordinates as the others. Variant grids break visually when one item is 5% larger or sits lower than its siblings.
Background color is exactly what the marketplace requires
Amazon and Walmart want exact RGB (255, 255, 255). Some removers output #fafafa or have slight blue tint. Use an eyedropper tool on the background pixel directly — eyeballing is unreliable.
Product fills the required product-area percentage
Amazon requires ≥85%, Walmart ≥75%. Measure by eye or against a consistent template overlay during QA. Small-object listings fail this check most often.
No watermarks or text bled through from the original
If the original had a timestamp, a label, or seller branding, the remover should have cleaned it up. Spot-check for residual text or logos — they usually show up in corners where the remover preserved background by mistake.
File is exported in the correct format and dimensions
The very last check is the easiest to skip: right format (JPG for Amazon, WebP OK for Shopify), right size (marketplace spec), right color profile (sRGB). An image that passes all other checks but exports as a 600 × 600 pixel PNG-24 still fails on the marketplace.

What to do when a check fails

Failed checks fall into three buckets: preset-level failures, photo-level failures, and product-level failures. The fix depends on which bucket.

Preset-level failures affect the entire batch uniformly. Common examples: wrong shadow direction, wrong product-area percentage, wrong background color. Fix the preset once and re-process the batch.

Photo-level failures affect a subset of the batch. Common examples: one photo had a different lighting condition, or one angle tripped the AI. Fix the individual photo manually or reshoot.

Product-level failures affect every photo of a specific product. Common examples: transparent product, reflective product, or fine-detail product. Switch to a specialized preset for that product category, or flag the product for manual editing.

Tools that help QA actually happen

QA only works if it's systematic. A checklist on a sticky note gets skipped the first busy week. The teams that maintain QA long-term make it part of the export pipeline, not a manual pre-upload step.

The practical setup is a hybrid pipeline: automate what you can validate quantitatively (dimensions, file format, background color), then sample human QA for perceptual issues like edge cleanliness, shadow quality, and color fidelity. That still cuts review time dramatically without pretending the whole problem is solved automatically.

Sampling protocol used by the teams shipping 10,000+ listings a month: automated pre-filter, then human QA on 10% of what passes, then ship. Approximate throughput: 900 listings per reviewer per day.

Frequently asked questions

How often should I run this QA checklist?

On every batch, but sampled — not every photo. Sample 10% minimum, 20% for categories with known AI weakness (hair, glass, reflective). For large batches (1000+ images) the sample is statistically sufficient; systematic errors will surface in the sample.

Which items in the checklist are most important?

Edge cleanliness, shadow direction, and background color match. Those three cover ~80% of buyer-perceivable quality failures. The other nine matter but are less frequently missed — edge and shadow are where the AI is most likely to fail subtly.

Can I automate this QA with AI?

Partially. Background color, resolution, file format, and other quantitative checks are good candidates for automation. The checks that still need human eyes are perceptual: does the shadow feel right, does the color look right, does the reflection make sense. Expect a hybrid pipeline — automated for quantitative, human for perceptual.

How do I fix a batch with systematic edge halos?

Halo usually means the remover's alpha threshold is too aggressive. In CatalogCut, adjust the edge feathering setting in the preset by +1 to +2 points. This softens the transition between product and background by a few pixels and removes the halo without losing product detail.

What's the biggest mistake sellers make on background removal?

Skipping sample QA. The preset looks right on the first photo, so they trust it for the whole batch. Then halfway through the batch, a photo with slightly different lighting produces a bad output, and it ships to the marketplace unchecked. A 30-second random sample of 5 images would have caught it.