AI-Generated Product Photos That Don't Look Fake: A Method

The fastest way to spot a fake product photo is to read it. Generated scenes are convincing now; generated label text still is not, and the label is the one part of the frame a shopper can check against the box that arrives at their door. This post applies our second-look standard to still images: a working method for AI generated product photos that starts from a real photograph, lets the model touch only what it is good at, and never regenerates the pixels that make legal claims.

A still image is a harder test than video

That sounds backwards — video is the harder generation problem — but the viewing conditions flip it. A video artifact exists for a tenth of a second and is gone. A product photo sits on a listing page while a shopper zooms in, compares it against the other six images in the gallery, and reads the ingredient panel checking for allergens. The second look isn’t something a skeptical viewer chooses to give a product photo; it’s how product photos are used.

That makes stills the purest case of the standard we defined in what is realistic AI: output a normal person accepts as real on the second look, not the first glance. For a product image, the second look happens at 200% zoom, and it happens every time someone is about to spend money.

Where AI generated product photos die

Three failure points account for nearly all the fakes that get caught.

Label text. Image models synthesize plausible texture, not exact glyphs. As of mid-2026, frontier models can often render a short brand name correctly; they still cannot reliably reproduce a dense ingredient panel, a barcode, a net-weight statement, or the trademark symbol your lawyer cares about. The same failure that makes background signage melt in video — item four on our detection checklist — is fatal in a still, because the resolution is higher and the viewing time is unlimited.

Reflective and transparent surfaces. A glass bottle refracts the scene behind it; chrome reflects the room it sits in. Generated reflections are plausible but incoherent: the bottle shows a window the shadows say isn’t there, or two products in one scene reflect two different rooms. Most shoppers can’t articulate what’s wrong. They just report that the photo feels off — which, for a purchase decision, is the same as wrong.

Contact and weight. Generated products float. The contact shadow where an object meets a surface — soft, dark, tight to the base — is subtle enough that models routinely fudge it, and its absence reads instantly as compositing. This is the same mistake bad Photoshop has made for thirty years; the tooling got smarter and the tell stayed identical.

The method: the real product stays real

The one decision that matters more than model choice, prompt style, or budget: start from a real photograph of the real product, and never let the model redraw it. This is the same cheat that anchors our video work — image-to-video from a real frame — applied one medium down. The model cannot misspell a label it never touched.

The input requirements are modest. A recent phone camera, the product near a large window with indirect light, the highest resolution your camera offers, several angles. You are not trying to shoot a beautiful photograph. You are capturing accurate product pixels for the machine to build a set around.

Three workflow tiers, from safest to riskiest

Workflow	What stays real	What’s generated	Label risk	Use it for
Real-photo base	Everything	Cleanup only: dust, color, background whitening	Near zero	Main listing images
Background swap	The product, behind a mask	The entire scene around it	Low, if the mask holds	Lifestyle and context shots
Full generation	Nothing	Everything, product included	High — assume the label is wrong	Mood imagery where the product is small or out of focus

Tier 1: real photo, AI cleanup

This is retouching with better tools — dust removal, white-balance correction, whitening the background to marketplace spec. The output is still a photograph of your product. It’s the only tier we’d put behind a main marketplace image without hesitation, and for a lot of sellers it’s also the only tier they actually need.

Tier 2: background replacement

The workhorse tier. Mask the product — most current tools do this in one click, though check the edges around handles, spouts, and hair-thin gaps — lock those pixels, and generate the scene around it: a kitchen counter, a bathroom shelf, a campsite table. Two disciplines keep it honest:

1. Verify the mask survived. Some pipelines quietly re-synthesize the entire image rather than compositing, and the “preserved” product comes back subtly rewritten, label included. After every generation step, overlay the output on your original at 200% and confirm the product pixels are identical — not similar, identical.
2. Match the light or reshoot. If the generated scene is lit from the left and your product was photographed lit from the right, no prompt fixes it. It’s cheaper to reshoot the product under neutral, diffuse light — which composites into almost anything — than to fight a mismatch. And add the contact shadow deliberately: ask the tool for it or paint it in. Floating products are the tier’s signature failure.

Tier 3: full generation

Sometimes fine — a blog header, a mood shot where the product is small, blurred, or facing away. The rule: full generation is acceptable exactly when the product’s specifics don’t matter, and a product listing is the definition of a place where specifics matter.

Marketplace rules: a product photo is a claim

We’re not going to quote specific platform policies, because they vary by marketplace, by category, and by quarter — any specifics printed here would be stale before you finished the post. The principle underneath them doesn’t move: a product image is a representation of the item the buyer receives, and every major marketplace holds sellers to some version of that accuracy standard, with main images held to stricter, more literal rules than lifestyle shots.

Follow the principle and the AI question mostly answers itself:

• A cleaned-up real photo makes the same claim a photo always made. Safe.
• A real product on a generated background claims the product precisely and the setting loosely — the same claim staged lifestyle photography has always made. Generally defensible; verify your category.
• A fully generated product claims an item that does not exist. When the render’s proportions, finish, or label differ from what ships — and they will — that’s a returns problem, then a suspension risk, and possibly an advertising-law problem, all independent of any AI-specific rule.

Some platforms are also adding synthetic-media labeling requirements, which stack on top of the accuracy rules rather than replacing them; the moving parts are mapped in our AI content disclosure guide. Two boring habits cover most of the risk: check the current image policy for your marketplace and category before uploading anything AI-touched, and archive your original photos as evidence that your images match reality.

The second-look pass for stills

Before an image ships, inspect it at 200% zoom with the physical product in hand:

• Read the label. Every character, including the small print. Not “does it look right” — read it.
• Check that reflections tell one story. Light sources implied by reflections should agree with the shadows and with every other object in the scene.
• Check contact. The product should sit, not hover. Look for the tight shadow at the base.
• Check geometry against the real item. Cap threads, seam lines, handle curves. Hold the product while you look.
• Check the set, not just the image. Every gallery image should show the same proportions and label layout. Drift across the set is a tell no single image reveals.
• Show it to whoever packs the boxes. They will spot a wrong proportion faster than any checklist.

Anything that snags: regenerate the scene, tighten the crop, or drop the image. The kill criteria are the same as in our video pipeline, minus the audio.

Where this loses, honestly

• Label reproduction inside generated regions. If a workflow requires the model to draw your label — an angle you never shot, say — expect it to be wrong as of mid-2026. Shoot the angle instead.
• Liquids in glass. Refraction bends the background scene through the product, so a background swap behind a glass bottle produces a physically impossible image unless the tool models refraction, and most don’t. Use a plain background you can live with, or a tighter crop.
• Highly reflective products. Jewelry and chrome mirror their environment, and a generated environment mostly fails to show up in those reflections. This is the category where a cheap lightbox beats any AI workflow on both time and quality. There’s no shame in equipment.
• Categories with literal-photo rules. Some categories — apparel on a model is the common example — carry image requirements a generated scene may fail regardless of quality. Category rules beat technique, every time.

None of that is an argument against the method. It’s the boundary of it, which is the part most guides skip.

The takeaway

Four moves, in order: shoot the real product in diffuse light at full resolution; tier your images — real-photo base for the main image, background swap for lifestyle shots, full generation only where the product is incidental; run the 200% pass on label, reflections, and contact before anything ships; and archive the originals while you check your marketplace’s current image policy.

AI generated product photos earn their keep when the AI is the set dresser, not the product photographer. The same second-look craft in motion is the harder version of this skill — how to make realistic AI videos documents that pipeline — and if you want it taught end-to-end, Realistic AI Club teaches photorealistic AI video and AI UGC as a repeatable craft for ten dollars a month.