Why Most AI Product Photos Look Fake, and What Fixes It

By the ORA Lab team · Updated 10 August 2026 · 9 min read

You've felt it even if you've never articulated it. An AI-generated product image where everything is technically in place, product centred, scene attractive, lighting dramatic, and something is still wrong. Not wrong enough to point at. Wrong the way a wax figure is wrong: the closer it gets, the more your brain objects.

Brands should care about this feeling because customers act on it without analysing it. An image that reads as fake quietly transfers that judgement to the product, synthetic photo, synthetic-feeling brand. The good news is that the uncanny effect isn't magic and isn't inevitable. It comes from a short, learnable list of physical errors, and each one has a specific cause and a specific fix.

Your eye is a physics auditor

Human vision evolved to model the physical world, and it audits images against physics constantly and unconsciously. You know where the light comes from in a scene without thinking about it. You expect a shiny surface to reflect its surroundings. You know how big linen weave is, how a shadow should soften with distance, how skin differs from plastic. When an image violates any of these, you don't get an error message, you get a feeling. 'Off.' That feeling is the sum of the violations, which means fixing the violations fixes the feeling.

The six tells, in order of how loudly they shout

1. Reflections that don't encode the scene

A polished ring in a real photograph is a tiny curved mirror: the window, the softbox, the room, they're all in the metal, distorted but legible. General AI models paint 'shininess' instead: bright smears that resemble reflections without corresponding to anything in the scene. On mirror-finish jewellery this is the single loudest tell, and it's why generic tools fail hardest on precisely the products where finish is the selling point.

2. Lighting disagreement

The scene says late-afternoon window light from the left; the product's highlights say studio softbox from above. Each element is plausibly lit, together they're impossible. This happens when the product is generated or composited with different implicit lighting than the environment, and it's the classic failure of naive cut-and-paste compositing as well.

3. Texture at the wrong scale

Wood grain sized for a floorboard on a jewellery box. Bouclé loops twice life-size on a chair. Skin pores absent entirely. Texture frequency is a strong subconscious size cue, and models frequently render material texture at whatever scale dominated their training data rather than the object's actual scale, making products read as miniatures or toys.

4. The plastic smoothness

Real surfaces are micro-imperfect: dust, faint scratches, weave irregularity, the soft asymmetry of anything hand-finished. Generative models regress toward clean averages, producing surfaces with no micro-history. The result reads as injection-moulded, fatal for categories whose value proposition is craft.

5. Shadow and contact errors

Objects meeting surfaces produce contact shadows, the dark seam that anchors a thing to the world. AI composites routinely float: no contact darkening, or a shadow direction that disagrees with the key light, or shadow softness that doesn't match the light source's size. A sofa hovering one millimetre above its floor is invisible consciously and glaring subconsciously.

6. Impossible geometry in the supporting cast

Background chair legs that don't reach the floor, window mullions that change spacing, a model's hand with ambiguous knuckles. Even when the product is perfect, scene errors leak distrust onto it. Buyers rarely inspect the background, but they absorb it.

Why models make these mistakes

Diffusion models learn what images of things tend to look like, not how light actually behaves. For most content, statistical resemblance is enough, skin, foliage, fabric in soft light. The tells cluster exactly where resemblance and physics diverge: specular reflection (which must encode this scene, not a typical one), global light consistency (which requires the whole image to agree), and scale-dependent texture (which requires knowing the object's true size). None of that is in a caption, so none of it is reliably learned. It has to be imposed.

What actually fixes it

• Domain training: models trained specifically on luxury product photography learn the material behaviours that matter, how brushed versus polished gold responds to a key light, what a bezel-set stone does with an edge highlight. This is the foundation of our research programme.
• Geometry and material conditioning: extracting the product's actual shape, size, and finish from the source photo and constraining generation to respect them, the same anchoring that prevents accuracy failures also produces scale-true texture and coherent form.
• Scene-consistent relighting: generating product shading from the scene's actual illumination rather than compositing pre-lit assets, one light model governing the whole frame kills tells #1, #2 and #5 together.
• Deliberate imperfection: reintroducing the micro-history real surfaces have, controlled grain, natural shadow softness, restrained highlight rolloff. Perfection is the artefact; a touch of entropy is the realism.
• Human review against the tell-list: thirty seconds per hero image checking reflections, light agreement, contact shadows, texture scale, and edges. The audit is fast once you know what to look for.

Auditing your own imagery: the five looks

1. 1.Reflection check, zoom into the shiniest surface: do the bright shapes correspond to anything the scene could contain?
2. 2.Light agreement, pick the scene's light direction from the shadows, then check the product's highlights agree.
3. 3.Contact check, does every object touching a surface have a contact shadow, and do all shadows point the same way?
4. 4.Texture ruler, is every material's texture scaled to the object's real size? (Wood grain and fabric weave are the usual offenders.)
5. 5.Edge quality, product silhouettes should be optically soft like a lens renders them, not cut-out crisp or smudged.

Run those five looks on any image before it ships and you'll catch nearly everything a customer would only feel. And if you want to see what scene-consistent, physics-respecting generation looks like on your own products, book a demo and bring your shiniest piece. Reflective products are where the difference is most visible, which is exactly why we like demoing with them.