Data report: We tested AI-generated Amazon product images from one packshot—30-second generation claims, listing-image compliance, and product-accuracy failure rates

In one observed run apiece, three Lamina workflows turned one packshot into an Amazon-style image within 30 seconds. That is all this test establishes. It does not show that any result met Amazon requirements, matched the product faithfully, or was fit to publish—and that gap is the whole report. Fast renders may help you make variants. No supplied measurement covered white-background compliance, frame fill, logo preservation, label accuracy, first-pass usability, retries, reviewer agreement, or Amazon acceptance.

The evidence supports one narrow operational claim: under these conditions, generation latency beat the test target. It does not justify the bigger vendor-style claim that one photo reliably produces a valid main image or a faithful multi-image listing set. Read this as a timing observation. It is not a certification test.

What did the one-packshot experiment actually prove?

The experiment proved one thing: all three tested workflows finished inside the 30-second usability target, with one observed run for each condition. Reference-lock finished fastest, the minimal baseline came next, and the compliance-constrained version took longest; each run cost $0.04 to generate. That is a modest draft-iteration budget. It is not a cost per approved listing image, since human review, revisions, retries, and publication work were left out.

One run does not give you a reliability distribution. With one observation per workflow, the median and p90 just repeat that observed latency; they tell you nothing dependable about a production batch. The central hypothesis has no pass/fail result either. Nothing supplied shows whether compliance prompting improved main-image compliance, raised fidelity failures, or whether reference-lock instructions reduced those failures.

Do 30-second claims mean you have a publishable Amazon image set?

No. A roughly 30-second render claim shows possible draft speed, not a publishable Amazon image set. Rawshot advertises roughly 30–40 seconds for each Amazon-style fashion generation, and Like.photo says a packshot can be finished in 30 seconds. Those are vendor capability claims, not independent tests of accuracy, approval, or full-set turnaround.

GreenOnion makes a similar claim: one uploaded product photo can produce nine listing-image types in 60 seconds. That could be a useful way to make candidate lifestyle, detail, and instructional concepts from a verified source image. The difficult work remains. Every generated view still needs to match the SKU, packaging, included items, and every factual claim before it goes into a listing.

Can AI-generated images meet Amazon main-image requirements?

AI-assisted edits can fit an Amazon main-image workflow if they preserve the actual product. A fabricated or materially altered product should not take the main slot. The supplied policy interpretations draw the boundary plainly: truthful background removal, white-background cleanup, and color or lighting correction may be acceptable, while an invented product depiction is described as non-compliant.

Treat the verified packshot as your source of truth. For the main image, stick to cleanup and compliant framing; do not ask a model to rebuild unseen product geometry, copy, or packaging. Put generated environments and concepts in secondary images, then review them at the item level. CNBC also reported that Amazon told sellers in July 2026 that AI-generated people in photos, videos, and A+ content must include specified metadata keywords. That disclosure requirement is separate from product accuracy.

What product-accuracy failure rate should ecommerce teams plan for?

Plan for intensive SKU review. In Photoroom’s benchmark, the strongest base model achieved full product fidelity in 29.0% of tested generations. Photoroom’s Fidelity Layer brought the reported pass rate to 38.2%, which still leaves a substantial share of outputs below that benchmark’s full-fidelity standard. These figures are neither Amazon rejection rates nor a measurement of Lamina. They are a warning: visual plausibility is not product truth.

Listings usually break on the small, costly details: buttons, zippers, logos, stitching, color, text, edge shape, scale, and packaging. Snappyit specifically says general image generators may create the look of letters instead of dependable words. One changed character on a label still makes a clean-looking image the wrong SKU representation.

What should ecommerce teams do with this data?

Use one-packshot generation for fast candidate images, then publish only outputs that clear documented fidelity and marketplace review. The timing result makes quick exploration plausible: all three tested drafts arrived within 30 seconds. Human art direction and approval remain the control point, especially for hero images, where a changed logo, dimension, or included accessory can mislead a buyer.

Do not call a workflow compliant or accurate until you measure it. A defensible next test would cover multiple products across categories, retain the exact source packshots and prompts, use independent reviewers to score compliance and product fidelity, track retries and operator time, and calculate cost per image that passes both checks. That is the evidence missing from a comparison between a fast draft workflow and a reliable publishing workflow.