Pro product photos from a phone (no studio)

A phone, a kitchen window, four AI tools, sixty-three cents in API credits. Forty-two minutes from a rough snapshot on a wooden counter to a six-shot product gallery good enough to upload to Amazon, Etsy, or your own Shopify store. No studio booking. No photographer. No second camera angle that nobody ever uses. This is the post that walks the whole pipeline, end to end, with every prompt and every dollar named.

The product for this run is a fictional thirty-five-dollar handmade ceramic mug from a one-person shop called Crater Goods. Pick your own: the recipe runs the same on a candle, a leather wallet, a face cream, a coffee bag, a piece of jewellery. The point is not the mug. The point is that a working seller, on a Sunday afternoon, can match a five-hundred-dollar studio invoice for a fraction of the time and money — if they treat AI as a craft, not a button.

The whole setup costs twelve dollars

Three things, in this order. A window that faces north or east, because north-facing light is the most consistent across the day and east-facing light is the softest in the morning. A small tripod with a phone mount — the cheapest one on Amazon will do. A piece of white foam board, two dollars at any craft store, used as a bounce card to fill the shadow side of the product.

Turn the iPhone’s grid on (Settings → Camera → Grid) and never use the flash. The HDR setting can stay on for textured products like ceramics, leather, and fabric; turn it off for shiny surfaces like glass and chrome where it tends to flatten highlights. Tape the product’s footprint to the counter so it ends up in the same spot for every shot — consistency in the source pays back twice in the AI step.

The source photo is a contract, not a portfolio piece

Six frames is enough. One from a 3/4 angle, one straight-on, one from the top, one from the side, and two close-ups of the texture (the rim of the mug, the bottom mark, whatever the buyer would actually want to see in zoom). The whole capture step takes eight minutes if you don’t fuss with composition. Don’t fuss. The AI step is what makes it look polished; your job is sharpness, even light, and a clean view of the actual product.

Two things will sink the AI pass if you skip them. Camera shake — without a tripod or a wall to brace against, even a sub-second blur gives the model permission to invent details. And mixed light — if you have an overhead bulb on plus the window, the colour cast shifts and the white-balance becomes a problem you cannot fix later. Turn off the bulb. Window only.

Background cleanup is a thirty-second step

Before the AI image step, knock out the background of every source shot. Photoroom’s free tier handles 250 background-remove exports per month at no cost (with a watermark on free downloads; Pro is $7.50/month annual to drop it). Drop in the photo, wait two seconds, download the transparent-background PNG. Native iOS does this too — long-press the product, “Copy Subject” — if you prefer to stay in the camera roll.

Why clean up before generating, not after? Because the next step uses your photo as a reference, and the AI model will treat anything in the frame as “the scene.” If you leave a coffee ring on the counter behind the mug, the model will faithfully reproduce a coffee ring on a hundred-dollar marble surface. The reference is a contract. The cleaner you give it to the model, the more control you keep over what gets generated.

Catalog shots: fidelity, not creativity

The hero shot — product on pure-white — is the one Amazon cares about most. Amazon’s 2026 spec is RGB 255,255,255 on the longest side at 1600px minimum (2000×2000 is the practical sweet spot), product filling at least 85% of the frame, no text or logos. The AI step is what gets your phone snapshot to that.

Three image models are credible in May 2026 for product fidelity. The criterion is narrow: does it preserve the actual product when you give it a reference? Black Forest Labs’ FLUX.1 Kontext Pro is the standout here. It is a context-aware editor, not a from-scratch generator, which means it already “sees” your source photo and you only have to describe what to change. At $0.04 per image on the Black Forest Labs API, a hero shot lands at four cents. Google’s Imagen 4 Fast at $0.02 per image is the cheapest credible option and runs at half the cost when you are iterating; Imagen 4 Standard at $0.04 is the keeper.

Here is the prompt that did the work for the white-background hero, run through FLUX.1 Kontext Pro with the cleaned mug photo as the reference:

Place the mug on a pure white seamless background, RGB 255,255,255.
Studio softbox lighting from the upper left, soft shadow on the right.
Keep the mug's exact silhouette, colour, glaze texture, and bottom mark
unchanged. Square 1:1 crop, mug filling 85% of the frame.
Do not add steam, liquid, props, or text.

Two failures in three generations, for honesty. The first pass invented a faint blue tint in the shadow because the prompt did not pin the background to pure white explicitly. The second pass moved the bottom mark a quarter-inch off-centre. The fix in both cases was a single added sentence: “keep the bottom mark in its exact original position.” Telling Kontext what to preserve is as important as telling it what to change. Quote the parts that must not move.

Lifestyle shots are where AI earns its keep

Three lifestyle scenes — a sunlit kitchen counter, a writing desk, a bedside table at golden hour — cover ninety percent of what a Shopify gallery needs. This is the part of the workflow that would have cost a studio day; with AI it is ten minutes and twelve cents on Imagen 4 Standard. Same source photo, different scene prompts. Generate three or four variants per scene and keep the one that ages best on a phone screen.

The discipline that matters here is the “keep” line. The model will happily redesign your mug to fit a more photogenic composition. Tell it not to. Lock the silhouette, the colour, the proportions; let it move freely with the props, the surface, the light, the background. Here is the kitchen-counter prompt that earned its place in the final gallery:

Place the mug on a warm oak kitchen counter, soft morning window light
from the left, a folded linen napkin and a small ceramic spoon-rest
behind it, slightly blurred background. Shallow depth of field, the
mug in sharp focus. Keep the mug's exact silhouette, colour, glaze
texture, and bottom mark unchanged from the reference.

Two patterns earn their keep. First: name the prop pieces precisely (“folded linen napkin” beats “napkin”) because a vague noun in the prompt becomes a generic noun in the frame. Second: name the light direction explicitly. “Window light from the left” reads in the final image; “soft natural lighting” collapses to a flat ambient and looks like every other AI shot. The specificity is what separates a real-feeling lifestyle scene from a stock-looking one.

On-model and texture: the brittle pillar

For wearable products — clothing, jewellery, hats, eyewear — you need an on-body shot. This is the part of the AI workflow most likely to embarrass you in 2026. The current best-in-class tools for virtual try-on (IDM-VTON for clothing, FLUX PuLID for character consistency) are credible at a glance and break at 200% zoom. Stitching, button placement, neckline shape, and small logos are the things that go wrong, and they are the things buyers look at.

The honest recommendation is to use on-model AI for the secondary gallery shots only, never for the main image, and to budget a human eye for a final review. If you are selling a $400 leather jacket, hire a model for the hero and use AI for the “styled on a chair” context shots. If you are selling a $14 cotton tee, AI on-model is probably fine for the whole gallery — the dollar value of getting it wrong is bounded. The trade is real; name it for your category.

For texture-heavy products — fabric weave, wood grain, glass, metal — do not skip the upscale pass. Topaz Gigapixel at $149/year for the Personal tier is the workhorse here; it runs locally on your machine, which means you can upscale a thousand images for the same flat fee. Magnific is the heavier creative upscaler at $39/month for the Pro tier and earns its place if you want it to invent texture detail on top of a slightly soft generation. Use Topaz for fidelity, Magnific for rescue jobs.

Marketplace policy: plain English, with dates

Three platforms, three different postures. Get this wrong and your listing gets suppressed on a Tuesday morning. Get it right and nothing happens, which is the goal.

Amazon permits AI-assisted images so long as they accurately represent the physical product shipping to the customer. Background swaps, lifestyle context, light adjustments, and props are all allowed. What is not allowed: generating features the product does not actually have (a stitched logo you have not embroidered, a strap that does not exist), misrepresenting scale or material, or fabricating before/after comparisons. Disclosure is recommended near the start of the product description. Initial violations trigger listing suppression; repeat offences can suspend the account.

Etsy is stricter on disclosure and softer on AI use. AI-assisted products must be listed under “Designed by” rather than “Made by”, and at least one image in the gallery must be a real photograph of the actual physical item. AI-generated context shots are fine alongside that real image; an all-AI gallery violates the authenticity requirement. Disclosure goes in the listing description, not the title.

Shopifyhas no AI-image restrictions at the platform level — Shopify Magic itself ships AI image tools and the App Store features dozens more. The catch is downstream: if you sell into the EU, the EU AI Act Article 50 transparency obligations come into effect on 2 August 2026, and AI-generated imagery that could be mistaken for a real photograph requires a buyer-facing AI label. US-only sellers are clear; international sellers should add a small AI-disclosure line to the product page now and forget about it.

The receipt

The whole pipeline takes forty-two minutes and costs sixty-three cents. Compare that to a traditional studio quote for a six-shot product gallery in 2026: $500 at the cheap end, $2,500 for the kind of shop that gets a hero shot right, plus a week of calendar time. Even if you mistrust the AI numbers by an order of magnitude — suppose the median seller actually spends two hours and ten dollars — the comparison still ends in the same place. The AI workflow is what makes a six-product Etsy shop possible on a one-person budget, and it is what makes a forty-product Shopify catalog cheap enough to refresh seasonally instead of annually.

Two honest caveats. First: the API costs assume you have a credit-card balance on each provider. If you start cold, factor in fifteen minutes to sign up and add a payment method on Black Forest Labs and Google Cloud, both of which require you to enable billing. Second: the forty-two-minute number is steady-state. Your first run will take longer — ninety minutes, perhaps, while you learn which prompt phrasings the models respect. By your fifth product, you will be at the forty-two-minute floor.

The pattern that travelled best out of this whole exercise is the same one that rescued the video-ad workflow last week: write the brief first, capture the source second, let the AI handle only the work that the source has constrained. For the broader argument about why a tight five-tool stack beats a sprawling subscription pile, the curation post is the place to start. And for the underlying read on which image models earned their keep this quarter, the spring 2026 image catalog runs the longer comparison. The tools will be different in 2027. The pipeline shape — phone, window, clean, generate, upscale, ship — will not.