AI product photos — when they work, when they fail.

Where AI product photos work on the first try

1. Packaged consumer goods

Boxes, cartons, bottles with clear labels. The model reads packaging geometry cleanly from the reference photo and renders the box with plausible materiality — cardboard, high-gloss laminate, matte foil. Labelling carries through reliably as long as it's legible in the reference.

Typical fail rate: under 10%. Usually shadow shape or background tone — both fixable in seconds.

2. Promotional goods with a fixed shape

Pens, mugs, keychains, USB sticks. Simple geometry, clean silhouette, low material variance. The models have seen thousands of these objects and reproduce proportions accurately. Logo prints from the reference carry over well.

Exception: fine engravings on metal. Here the model loses detail. The workaround is a separate detail crop as a secondary image.

3. Household and office goods

Binders, bins, coffee machines, desk organisers. Medium complexity but well inside the training distribution. They work very well in studio packshots and acceptably in lifestyle scenes.

4. Furniture and larger one-off pieces

Sofas, shelving units, office chairs. As long as the base shape is clear, the model renders these objects convincingly in interior scenes. The catch: dimensions and proportions have to be readable from the reference, otherwise scale drifts.

Where AI product photos consistently break

1. Small metal parts with glossy surfaces

Screws, hand tools, hardware, fittings. The problem isn't the objects — it's the reflections. Chrome surfaces reflect the studio environment, and the model invents that environment. The result: geometrically correct objects with reflections that don't match the rendered setting.

Workaround so far: light retouching on the reference photo, not a full re-render.

2. Patterned textiles

Patterned fabrics, weaves, specific prints. The model renders "a similar fabric", not "this fabric". For catalog images that need to show the pattern exactly, this path fails.

Escape hatch: solid-colour textiles work; patterned ones need real photos.

3. Food and drink with colour fidelity

Wines, spirits, juices, fresh produce. The model hits the category but not the specific colour tone. A rosé looks like "a rosé", not like the customer's specific 2022 vintage.

For brands where colour is part of product identity, AI generation isn't the right tool.

The 60/40 rule

Across every category we've tested, the "pass on the first try" rate sits around 60%. That sounds mediocre, but with automated regeneration it's easy to handle — one fail/regen round brings the combined rate to about 88%, two rounds to 95%.

What matters is that the review gate exists. Without it, the 40% of bad images land in the shop — and the project fails, not the technology.

A pragmatic decision rule

If at least 80% of your assortment falls into "packaging", "promotional goods", "office/household", or "furniture", the AI path pays for itself. The remaining 20% can be filled with manufacturer imagery or a targeted small studio session for the problem categories.

The mistake is deciding everything either/or. Hybrid — AI for the bulk, studio for the exceptions — is almost always the cheapest and fastest option.