Prompt Engineering for Product Images That Brands Can Actually Use

Generating a pretty bottle on a marble slab is easy. Generating an image a brand can legally publish, that shows their bottle, with their label, in a scene that matches their guidelines, is a different job. The gap between those two things is where most AI product work quietly fails.

Reference-first, not text-first

For a real product, text prompts alone are the wrong starting point. The workflow that survives client review begins with a clean photograph of the actual product — even a phone shot on a white sheet — and uses generation to build the scene around it rather than to invent the product. Whatever the tooling, the rule holds: the shape, proportions and label of the physical item come from reality; the environment, lighting and mood come from the model.

Where a client has no product photo yet (a concept, a pitch deck, a pre-production sample), I say so on the delivery and label those images as concept visuals. Presenting a generated approximation as a real product photograph is how studios lose clients.

Write lighting the way a photographer would

Most weak product prompts describe the object and forget the light — which is the actual subject of product photography. I write light explicitly, in four parts:

• Key: "large softbox at 45 degrees, camera left, soft shadow falloff".
• Fill: "white bounce card camera right, shadows lifted but not flat".
• Rim or kicker: "narrow strip light behind, separating the bottle from the background".
• Surface behaviour: "reflective glass with a single clean highlight strip; no double reflections".

Add lens language — focal length, aperture feel, camera height relative to the product — and you get repeatable results instead of lucky ones. A 100mm at eye level says "catalogue"; a 35mm slightly below says "lifestyle".

Build a scene block you can reuse

Brands buy sets, not single images. I keep a scene block per campaign — surface material, backdrop colour, props, palette, time of day — and change only the camera angle between shots. The result is a coherent set of six to twelve visuals that look like one shoot, which is what a content calendar actually needs. Changing three variables at once produces twelve orphans.

Labels, text and the accuracy problem

Generated typography is still unreliable, and a misspelled brand name on a hero image is unforgivable. My rule: never let the model write the label. Either composite the real label in post, or shoot the angle so the label is legible from the source photograph. For claims — "SPF 50", "500ml", certification marks — verify against the client's approved artwork, never against what looks right.

Brand safety review before delivery

Every set goes through a short review before it leaves the studio:

• Does the product match the physical item in shape, colour and finish?
• Is the label accurate and legible, with no invented text anywhere in frame?
• Do the props and setting fit the brand's audience and any cultural or religious sensitivities in the target market?
• Are there people in frame? If yes, are they clearly non-identifiable and disclosed as AI-generated?
• Any accidental logos, competitor cues or copyrighted artwork in the background?
• Does the image imply a claim the product cannot make — medical results, ingredients, endorsements?

Deliver in the shapes the client will use

I generate at the widest useful framing and crop out, delivering square, four-by-five and nine-by-sixteen versions of the same hero, plus a version with clean negative space for copy. Marketing teams re-crop badly; giving them intentional crops protects the composition and saves a round of revisions.

Where AI genuinely wins here

Not in replacing a photographer for a flagship campaign, but in volume and iteration: seasonal variants, background swaps for different markets, a dozen social crops from one concept, and rapid concepting before anyone books a studio. Treat it as an extension of the art department, keep the physical product as the anchor of truth, and the output stops looking like AI and starts looking like a shoot.