Product photography for AI: how to shoot images a model can work from

Generating a product into a new scene is a compositing problem wearing a machine-learning hat. The model has to understand the object's shape, its material and where its edges are, and then relight it for a setting that did not exist when you took the photograph.

Everything that makes that easy happens at the shoot, and almost none of it is expensive. A phone on a tripod with two windows will beat a studio photo shot the wrong way.

What to shoot

For each product you intend to generate from, capture a small, boring set:

1. A clean three-quarter view on a plain, mid-grey background. This is the workhorse and most generations will use it.
2. A straight-on front, for anything where the label or the front face matters.
3. A detail shot of texture — fabric weave, glass thickness, grain — so material is unambiguous.
4. One with a familiar object beside it, for scale. Models get size badly wrong without a reference, and a 30 ml bottle rendered as a litre is an obvious fake.

Mid-grey rather than white: pure white blows out the edges of light products and destroys the very thing the model needs.

Light it flat and even, not dramatically

Dramatic lighting is baked into the pixels and fights every scene you later place the product into. A hard shadow from the left cannot be relit to match a window on the right.

• Two soft sources, roughly 45° either side, or one big window and a white card.
• No coloured gels, no lens flare, no reflections of the room in glossy surfaces.
• Keep it slightly underexposed rather than clipped. Detail can be recovered from shadows; it cannot be invented in blown highlights.
• Shoot at f/8 or so. A shallow depth of field looks good on its own and gives the model a soft edge to guess at.

The four mistakes

1. Busy or coloured backgrounds. Colour bounces onto the product and follows it into every generated scene.
2. Existing shadows on the surface. They pin the product to a table that is about to be replaced.
3. Motion or hand-held blur. A model sharpens what it can and invents the rest, usually on the label.
4. Heavy retouching before generation. Smoothed texture reads as plastic once relit. Retouch after, not before.

What to check before the image goes out

The generation is a draft. A person approves it, and these are the four things to look at, in this order:

• The label and any text. This is where generation fails most visibly and most often. Text that is nearly right is worse than text that is obviously wrong.
• Proportion against surroundings. Compare to a known object in the frame.
• Shadow direction agreeing with the light in the scene.
• Material behaviour — glass refracting, metal reflecting, fabric creasing where a real one would.

If your workflow cannot show the original and the generation side by side, you will approve things you would have rejected.

What not to generate at all

Two limits worth writing into your own guidelines rather than discovering in a complaint. Never generate the product itself — the shape, the colour, the packaging must come from the real photograph, because a generated variation is a picture of something you do not sell. And never generate a claim: a dish the food does not look like, a result the cream does not produce, a size the garment is not.

The background can be invented. The product cannot.