AI product photography: the same perfume bottle, shot by 6 models

Nov 10, 2026 · 8 min read · By the Pizyn team

A product shot is the hardest brief you can hand an image model: glass has to refract, a label has to spell the brand right, and the light has to look like a studio, not a render. So we gave six models the exact same perfume-bottle prompt, kept every first result — no re-rolls — and priced each shot. The cheapest image in the set costs 1 credit. The most expensive costs 24. Both spelled the label correctly.

Amber glass perfume bottle with a LUMIÈRE label on a travertine pedestal, generated by FLUX 2 Pro

How the test ran

On November 10, 2026 we ran one text-to-image generation per model in this studio, all at 1:1, all from the same prompt, and published the first output of each — the good and the flawed. Settings match what the studio charges: the resolution and quality tier in each caption is the one the credit price buys. Nothing here is a press image or a vendor sample; every picture on this page was generated for this article, and the prices are the live studio prices from our pricing page.

"Studio product photograph of a minimalist amber glass perfume bottle with a matte black cap, label reading 'LUMIÈRE' in elegant serif letters, standing on a warm beige stone pedestal, soft diffused window light, gentle shadow, warm cream background, photorealistic, high-end commercial photography"

Why a perfume bottle? It compresses four failure modes into one frame: transparent material, small text with an accented capital (È), a reflective surface, and a pedestal the model has to keep physically plausible. If a model can shoot this, it can shoot most e-commerce listings. Our model guide compares all fourteen models by job; this is the product-photography chapter, run for real.

The shootout: six models, one prompt

Perfume bottle with LUMIÈRE printed directly on the glass, generated by Z-Image Turbo — **Z-Image Turbo · 1 credit**The 1-credit draft, and the surprise of the set. Correct spelling, real window light, a believable stone pedestal. It printed the name straight onto the glass instead of a paper label, and the bottle runs a little tall — but for a first composition test, this is absurd value.

Perfume bottle with a white paper LUMIÈRE label on a travertine pedestal, generated by FLUX 2 Pro — **FLUX 2 Pro · 5 credits**The catalog shot. A proper paper label with correct serif spelling, deep amber liquid, hard warm sunlight and a long shadow across travertine. The most "buy this" image of the six, at the lowest flagship price.

Frosted amber perfume bottle with a framed LUMIÈRE label, generated by Seedream 5 Pro — **Seedream 5 Pro · 1K · 8 credits**Frosted glass, a thin framed label, spelling correct. The softest light of the six — even and flattering, but the lowest contrast. The pedestal also crowds the frame; Seedream read "standing on a pedestal" as the composition, not the product.

Perfume bottle with a LUMIÈRE label plus invented EAU DE PARFUM text, generated by GPT Image 2.6 — **GPT Image 2.6 · Low · 5 credits**The most designed label — and the one that invents. It added "EAU DE PARFUM / 50 ML e 1.7 FL.OZ." under the brand name, all spelled correctly, with window-blind shadows across the wall. Charming for a fictional brand; check every invented line before a real SKU ships.

Perfume bottle with a clean white LUMIÈRE label in soft directional light, generated by Qwen Image 3.0 Pro — **Qwen Image 3.0 Pro · 1K · 7 credits**The quiet one. Clean white label, correct spelling, soft directional light, balanced exposure, nothing overcooked. Less drama than FLUX, more restraint — the shot you pick when the listing page, not the bottle, should do the shouting.

Cylindrical amber perfume bottle with a white LUMIÈRE label in window light, generated by Nano Banana Pro — **Nano Banana Pro · 2K · 24 credits**The only 2048px image here, and the most photographic: window light, shallow depth, travertine you can feel. But note the shape drift — every other model drew a rectangular bottle; Nano Banana Pro made it cylindrical. If the bottle has to match a real product, that is a fail, not a flourish.

What it cost

The six shots plus the edit below cost $0.317 in provider fees — 55 studio credits in total, about $0.55 at pack prices. The spread is the story: the 1-credit draft and the 5-credit FLUX shot together cost less than a third of the 24-credit Nano Banana Pro frame, and for a listing thumbnail nobody would know which was which without the caption.

Z-Image Turbo — 1 credit ($0.005). Drafts, composition tests, twenty variations of an idea. Our Z-Image Turbo page has the details.
FLUX 2 Pro — 5 credits ($0.03). The value pick for final product shots. See FLUX 2 Pro.
GPT Image 2.6 — 5 credits at Low ($0.03). The label-design pick, with the invented-copy caveat. See GPT Image 2.6.
Qwen Image 3.0 Pro — 7 credits at 1K ($0.04). The restrained, high-fidelity pick. See Qwen Image 3.0 Pro.
Seedream 5 Pro — 8 credits at 1K ($0.045). The soft-light pick. See Seedream 5 Pro.
Nano Banana Pro — 24 credits at 2K ($0.14). The resolution pick — and the one to watch for shape drift. See Nano Banana Pro.

What the results actually say

Label spelling is no longer the gamble it was. All six models spelled "LUMIÈRE" correctly, accent included, on the first try. A year ago that sentence would have been the whole article. The remaining text risk is not spelling but invention: GPT Image 2.6 added plausible label copy nobody asked for. For a fictional brand that is a gift; for a real product, where the label is a legal document, every extra line is a liability. Check text at full zoom before anything ships.

Shape is the new failure mode. Five models drew the same rectangular bottle; Nano Banana Pro drew a cylinder. Nothing in the prompt specified the silhouette, so every model guessed — and they guessed differently. If the product already exists, do not generate it from text alone: upload a photo of the real bottle and edit the scene around it instead. Our editing guide covers reference-based workflows, and the models take up to 16 references in the studio.

Light is a style choice, not a quality score. FLUX went hard and warm, Seedream went soft and even, Qwen Pro went restrained. All three are defensible product photography; which one fits depends on the brand. The prompt asked for "soft diffused window light" and every model obeyed it differently — prompt words set a direction, the model picks the weather.

From shoot to listing: the 5-credit background swap

A shoot is rarely one image. Listings need the same product on a cream background, a sage background, a seasonal background — and re-shooting the bottle six times is how budgets die. So we took the FLUX winner and asked Seedream 5 Flash to change only the scene: sage green backdrop, light grey marble pedestal, bottle untouched. One edit, 5 credits, same price as a generation — Flash charges the same in both modes, references included.

Original FLUX 2 Pro perfume shot on a warm cream background — **Before — FLUX 2 Pro, 5 credits.**The shootout winner, warm cream background, travertine pedestal.

Same perfume bottle on a sage green background with a grey marble pedestal, edited by Seedream 5 Flash — **After — Seedream 5 Flash edit, 5 credits.**Sage green backdrop, grey marble pedestal. The bottle, cap and label survived the swap untouched — including the spelling.

The edit prompt was explicit about what not to change — "keep the perfume bottle, its label and the lighting exactly the same" — and that is the whole technique. Editing prompts should spend most of their words on what stays put. Note the pricing asymmetry this exposes: GPT Image 2.6 edits start at 8 credits and add 3 per extra reference, and Qwen edits start at 6 — so for background swaps, a flat-rate editor like Flash is the cheap seat. Our editing guide lists every model's edit price.

The workflow we would actually use

Draft on Z-Image Turbo. At 1 credit, run ten variations of the composition — pedestal height, bottle angle, light direction — and pick a direction before spending anything real. This is the same draft-cheap, finish-expensive workflow from our prompt guide, and product work is where it pays off fastest.

Finish on FLUX 2 Pro or Qwen Image 3.0 Pro. FLUX for the warm catalog look at 5 credits; Qwen Pro for restrained, high-fidelity listings at 7. Reach for Nano Banana Pro only when you need the 2K or 4K file — and check the bottle shape against the real product before paying 24 credits for it.

Variant on Seedream 5 Flash. One winning frame, edited into every background your store needs, at 5 credits a variant. A six-background listing set costs 30 credits of edits on top of the one hero shot — cheaper than six fresh generations on any flagship.

Writing the prompt: the anatomy that worked

The shootout prompt is built in the order models seem to weight most: subject first, then material, then text, then staging, then light, then medium. "Minimalist amber glass perfume bottle with a matte black cap" fixes the object; "label reading 'LUMIÈRE' in elegant serif letters" puts the exact text in quotes, which our prompt guide recommends for anything that must be spelled right; "standing on a warm beige stone pedestal" stages it; "soft diffused window light, gentle shadow, warm cream background" lights it; "photorealistic, high-end commercial photography" sets the register. What it does not do is specify the bottle's silhouette — and that omission is exactly why five bottles are rectangular and one is cylindrical. For a real SKU, add the shape ("rectangular bottle, flat shoulders") or, better, upload the product photo and edit around it.

The short version

Best value product shot: FLUX 2 Pro, 5 credits — catalog-ready light and a correct label.
Best draft: Z-Image Turbo, 1 credit — correct spelling, real window light, nearly free.
Best label design: GPT Image 2.6 Low, 5 credits — but audit every line it invents.
Best restrained listing: Qwen Image 3.0 Pro, 7 credits at 1K.
Best background variants: Seedream 5 Flash edits, 5 credits flat, bottle preserved.
Watch out for: shape drift on text-to-image product shots — generate from a reference photo when the product already exists.

Every price above is the live studio price, and credits work across all fourteen models — see pricing for the packs. New accounts start with 20 free credits, which is enough to run this whole shootout yourself: six shots and a background swap cost 55 credits, and the draft round costs one.

Frequently asked questions

Can AI generate product photos for a real product I sell?

For the scene around it, yes; for the product itself, generate from a reference photo, not from text. In this shootout, five models drew a rectangular bottle and one drew a cylinder from the same prompt. Upload your product photo and edit the background, lighting or setting instead — the shape, label and proportions stay true.

Which AI model is best for product photography?

For a final listing shot, FLUX 2 Pro at 5 credits was the best value in our test: correct label spelling, catalog-style light and a believable pedestal. Qwen Image 3.0 Pro (7 credits at 1K) is the more restrained alternative, and Nano Banana Pro (24 credits at 2K) is the pick when you need a 2048px file — with a check on bottle shape first.

How much does an AI product photo cost?

From 1 credit for a Z-Image Turbo draft to 24 credits for a 2K Nano Banana Pro final. Our six-shot comparison plus a background swap cost 55 credits in total, about $0.55 at pack prices. Background variants via Seedream 5 Flash edits cost 5 credits each.

Will AI misspell my brand name on the label?

Less often than it used to: all six models in this test spelled "LUMIÈRE" correctly on the first try. The current risk is invented copy — GPT Image 2.6 added plausible label lines nobody asked for. Always check label text at full zoom, and put the exact brand text in quotes in your prompt.

About the author

Written by the Pizyn Studio team. We build and operate this 14-model AI image studio, and every example in our guides was generated in the studio itself, with the exact prompts shown and at the prices listed. Prices and model details are checked against the live studio on the date above. Spot an error? Email support@pizyn.com and we will correct it.

Try the models in this guide

FLUX 2 Pro · Seedream 5 Pro · GPT Image 2.6 · Qwen Image 3.0 Pro · Nano Banana Pro · Seedream 5 Flash · Z-Image Turbo · All 14 models