A perfume bottle is one of the most punishing things to photograph in a real studio. It is heavy faceted glass that mirrors the whole room and magnifies whatever is behind it, filled with colored juice that transmits and reflects light, topped with a polished metal cap. Getting it right means hours of careful lighting and gradients to control glare and reveal the facets. It is exactly the kind of subject you would expect AI to mangle.
It did not. I ran one brief, a heavy faceted clear-glass flacon of amber fragrance with a gold cap on a cream stone surface, refraction through the thick glass and subtle caustics, no logos, through four of the strongest image models with the same prompt: Nano Banana 2, GPT Image 2, Seedream 4.5, and FLUX.2 Pro. All four rendered the glass and juice beautifully, none cloned a famous bottle, and one made refraction the others could not touch. This is the perfume entry in our product-photography series, alongside the skincare, jewelry, supplements, makeup, food and beverage, footwear, candles, clothing, furniture, electronics, handbags, sunglasses, glassware, flowers, watches, packaging, pet products, toys, textiles, cookware, stationery, drinkware, soap, ceramics, art prints, earbuds, houseplants, knives, and automotive wheels tests and the broader best AI image model for product photography roundup.
Quick answer
- Best overall, and cheapest photoreal: Seedream 4.5. The most exquisite faceted, prismatic refraction of the juice, a premium macro.
- Glass and juice are solved: all four rendered believable thick-glass refraction, juice, caps, and caustics. The feared transparency problem did not appear.
- No brand clone: unlike the smartwatch and watch tests, none of the models copied a famous flacon. They invented generic bottles.
If you only remember one thing: perfume is a category AI handles beautifully, especially Seedream. The only real cautions are the exact bottle shape and any fine etched branding.
The test, model by model
One brief, four models, same prompt. I judged the thick-glass refraction first, then the juice and cap, then whether any model cloned a famous design.
Seedream 4.5 made the most beautiful flacon by a distance, and perfume is the category where its strength pays off cleanly. Across this series Seedream wins on material and optics but stumbles on fine text. A brand-free luxury bottle is all material and optics and no fine text, so it plays entirely to its strength: the many facets each refract the amber juice differently into a prismatic, kaleidoscopic effect, the caustic warms the marble, and the gold cap is detailed. Best result, lowest cost of the photoreal options.
GPT Image 2 produced the cleanest, most premium rectangular flacon. The beveled clear-glass edges magnify the amber juice the way real thick glass does, and the raking light throws a crisp caustic onto the surface, the same caustic strength it showed on glassware. It is a polished, catalog-ready result on a generic design, at the highest price of the four.
Nano Banana 2 gave the most balanced, finished shot: the faceted glass refracts the amber juice convincingly, the cap is clean, and the bottle rests on a soft surface reflection. It is not the prismatic showpiece Seedream produced, but as a ready-to-list hero with believable thick-glass optics and an accurate juice color, it is excellent and a third of GPT's cost.
FLUX.2 Pro leaned into the light. Its thick glass refracts the juice well, and it cast the brightest, most directional amber caustic of the four onto the marble, a genuinely cinematic light-driven shot. The overall rendering is a touch softer than Seedream's or GPT's macro, the usual FLUX tradeoff, but at the lowest cost it is a strong hero where the light is the point.
The comparison
| Model | Thick-glass refraction | Juice + cap | Cloned a bottle? | Rough cost/image |
|---|---|---|---|---|
| Seedream 4.5 | Best, prismatic faceted | Rich, detailed cap | No, generic | ~4.8 credits |
| GPT Image 2 | Excellent, clean bevel | Rich, clean cap | No, generic | ~26.4 credits |
| Nano Banana 2 | Strong, balanced | Accurate | No, generic | ~9.3 credits |
| FLUX.2 Pro | Good, strongest caustic | Good, softer | No, generic | ~3.6 credits |
Credit costs are first-hand from this test on Masonry; per-image rates move, so check current pricing.
Why perfume is the category that suits AI best
After a run of tests where the model broke something, a label, a screen, a dial, perfume is a high point, and it is worth understanding why.
It is all optics and no fine text. A brand-free luxury flacon is exactly the subject these models are best at: thick glass, refraction, colored liquid, metal, and caustics, with no dense printed information to garble. That is why Seedream, which stumbles on small labels and watch sub-dials, is the runaway winner here. Remove the text problem and its material strength has nothing to trip on.
The trade-dress risk did not appear. Unlike the Apple Watch and the TAG-Carrera watch, no model cloned a famous flacon. Iconic perfume bottles exist, but the design space is varied enough that the models invented generic shapes instead of defaulting to one. That is a genuine relief for this category, though it is still worth a glance.
The caveats are shape and branding. Each model produced a different bottle, so the exact silhouette is the model's invention, not your SKU, and for fragrance the bottle shape is often the trademark. And any fine etched or printed branding would face the usual AI-text weakness, which this brand-free test did not exercise. For your real bottle, generate from a reference and proof any branding.
How to shoot your fragrance line without a studio
The workflow is the roundup approach, applied to the subject AI handles best. For a brand-free or minimally branded bottle, trust the optics and choose on the refraction and the caustic, Seedream for the most exquisite, FLUX for the most dramatic light. For your exact flacon, feed a reference photo so the shape is yours, and if the bottle carries etched text, proof it at full size or composite the real branding.
With the Masonry CLI you can compare the refraction across models, or pass your real bottle as a reference to keep the exact design:
masonry image "heavy faceted glass perfume bottle, amber juice, gold cap, cream marble, soft light, refraction and caustics, photoreal" --model seedream-4-5 masonry image "place this exact perfume bottle on a sunlit marble surface, keep the shape and label" --ref ./real-bottle.png --model gpt-image-2
The bottom line
Perfume is the category that suits AI best in this whole series. The thick-glass refraction and the juice that make it a studio nightmare are exactly what these models render beautifully, and none of them cloned a famous bottle. Seedream 4.5 is the standout, the most exquisite faceted refraction at the lowest cost, with GPT Image 2 and Nano Banana 2 close behind and FLUX.2 Pro best for a light-driven shot. Use a reference photo for your exact bottle, proof any branding, and otherwise trust the optics. See how the same fidelity-first logic plays out across every product type in our best AI image model for product photography roundup, or run your own bottles from one place with the Masonry CLI.
