AI generation QA layer for photo studios

A photo editor at a commercial studio uses generative fill or AI erase on a batch of product shots. Some results look fine. Some are subtly wrong — a hand has six fingers, the background fill breaks at the edge, a reflection doesn't match the light source. The editor doesn't catch it until delivery, or worse, the client does.

The gap here isn't the generation quality itself — it's that there's no automated QA step between 'AI ran' and 'file goes out.' Every AI generation tool assumes a human will review every output manually. At any volume above a few images, that assumption breaks. The complaints about 'strange-looking' results, unrealistic hands and facial features, and generations that 'don't consistently achieve good results' are all things a trained classifier can flag automatically.

This gap persists because the generation vendors (Adobe, etc.) are incentivized to show high generation counts, not to surface their own failures. Adding a 'this looks wrong' flag would make their quality numbers look bad internally. A third party has no such conflict.

A studio without this either over-relies on manual spot-checks (slow, expensive, inconsistent) or ships bad AI fills occasionally and eats client revision costs. Neither is acceptable at volume. The need recurs every batch — and as studios lean more heavily on AI to hit throughput targets, the stakes go up.

What to build

Build a Lightroom Classic plugin and standalone batch processor that runs after generative AI edits are applied, scores each image for common AI artifacts — anatomical anomalies, edge discontinuities, lighting inconsistencies, text distortion — and flags low-confidence outputs for human review before export.

Where to start

Enter through event photography studios using AfterShoot for culling who already automate part of their workflow — they understand AI-assisted editing, already feel the pain of occasional bad outputs, and have a clear cost model where one bad delivery justifies the subscription.

The hard part

Training a classifier specific enough to catch subtle AI artifacts (a slightly wrong hand, a background fill that's almost right) without generating so many false positives that editors ignore the flags — the signal-to-noise ratio has to be right from day one or the tool gets turned off.

How it makes money

Per-seat monthly subscription for studio editors, with pricing tiers based on monthly image volume — similar to how proofing software charges — starting around $29/month for freelancers and scaling to $99+/month for studio teams.