A packaged-food box looks like the easy case. It is a cardboard rectangle with a nice illustration, and AI is good at boxes and illustrations. But a retail package carries two things that are not design, they are data: a barcode that has to scan, and a nutrition panel that has to be exact and is regulated. Those are the parts an image model cannot actually produce, and they are the parts that get a product rejected from a shelf or a marketplace.
So I tested it. I ran one brief, a brand-free artisan cracker box with a front illustration, a barcode, and a nutrition facts panel, through four of the strongest image models with the same prompt: Nano Banana 2, GPT Image 2, Seedream 4.5, and FLUX.2 Pro. The fronts came out great. The functional elements did not: every barcode is fake, every panel is invented, and the results split on something as basic as spelling the product name. This is the packaging 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, perfume, 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 front-of-pack: GPT Image 2. The most retail-ready box, legible copy, clean layout, and the most convincing (and therefore most dangerous) fake nutrition panel.
- The barcode is always fake: all four drew a barcode that will not scan. A barcode is data, not a texture. Never ship a generated one.
- The panel is always invented, and Seedream 4.5 garbled even the brand name into "Artisian cruckers," its usual fine-text weakness on package copy.
If you only remember one thing: AI makes a beautiful box, but the barcode and the nutrition panel are fake on every model. Composite the real ones, always.
The test, model by model
One brief, four models, same prompt. I judged the front-of-pack, then the two functional elements, the barcode and the panel, where packaging is won or lost.
GPT Image 2 produced the most convincing retail package. The front-of-pack copy is legible and well set, the illustration is professional, and the side carries a structured Nutrition Facts panel that looks entirely real. That is exactly the problem, the same one supplements surfaced: the strongest text model makes the most believable fake, a panel a busy team might actually ship. The better the text, the more carefully you have to stop it from writing your regulated copy.
Nano Banana 2 gave the most useful illustration of the core problem, because its barcode is clearly visible. The front copy is legible, the panel is structured, and at the bottom of the side panel is a barcode, a clean pattern of bars with digits beneath. It is also completely fake: the bars do not encode a number and the check digit is not valid, so it will not scan. This is the single clearest reason you cannot ship an AI package as-is.
FLUX.2 Pro surprised on the front: its minimalist "Artisan Oven" box has clean, legible front-of-pack copy, better than FLUX's usual incidental-text wobble, and a tidy design at the lowest cost. The side panel and barcode are present and, like every model here, fake. For a front-of-pack concept on a budget it is a strong result; for the functional side, it is the same composite-the-real-thing story.
Seedream 4.5 made the most beautiful image, a styled set of colorful illustrated boxes that looks like a paid packaging shoot, and then undid it on text. The brand name reads "Artisian cruckers," the panel header "Nutrilon Facts," the same garble it produced on skincare labels and watch sub-dials. For a mood or a styled scene it is stunning; for any shot where the package copy reads, it is unusable without compositing real text.
The comparison
| Model | Front-of-pack copy | Nutrition panel | Barcode | Rough cost/image |
|---|---|---|---|---|
| GPT Image 2 | Best, legible | Convincing but invented (risk) | Fake | ~26.4 credits |
| Nano Banana 2 | Legible | Structured, invented | Fake, clearly shown | ~9.3 credits |
| FLUX.2 Pro | Legible, clean | Present, invented | Fake | ~3.6 credits |
| Seedream 4.5 | Garbled ("Artisian") | Garbled ("Nutrilon Facts") | Fake | ~4.8 credits |
Credit costs are first-hand from this test on Masonry; per-image rates move, so check current pricing.
Why packaging is a data problem, not a design problem
For most products in this series the failure is a detail. For packaging the failures are the two things that make a package a sellable product, and both are data, not design.
A barcode is data, not a texture. A real barcode encodes a number with a valid check digit, and a scanner reads it. An image model draws bars because it has seen bars, but there is nothing behind them. Every model here produced a fake. That is not a quality difference between models, it is a category fact: you cannot generate a working barcode from a text prompt, you composite a real one. It is the most verifiable failure in this whole series, you can literally try to scan it.
The nutrition panel is the supplements problem again. A Nutrition Facts panel is regulated, exact information, and an image model invents it. GPT Image 2 made the most convincing one, which makes it the most dangerous, because it passes a glance and someone might trust it. Treat the panel like the regulated document it is and never let a model write it.
And the front copy is not guaranteed either. Brand name, flavor, weight, claims, these are dense text, and while GPT, Nano, and FLUX kept them legible, Seedream garbled the brand itself. So even the design side needs proofing. The illustration and the box will be great; the words are the risk.
How to shoot your packaging line without a studio
Use AI for what it is genuinely good at on packaging, the box structure, the illustration style, the materials, and the lifestyle scene, and protect every piece of text. Composite your real front-of-pack copy where exactness matters, and always composite the real barcode and the real nutrition panel rather than trusting generated ones. Better still, start from a flat of your actual artwork as a reference so the design is yours, and proof every word, then scan the barcode, before anything goes live.
With the Masonry CLI you can generate the box and scene across models, or pass your real artwork as a reference so the design stays yours and you only composite the regulated parts:
masonry image "artisan cracker box on a kitchen counter, soft morning light, retail packaging shot, photoreal" --model gpt-image-2 masonry image "place this exact box artwork on a sunlit shelf, keep the front design" --ref ./real-artwork.png --model gemini-3.1-flash-image-preview
The bottom line
Packaging is the category where the prettiest result is the most misleading. AI makes a beautiful box, GPT Image 2 the most retail-ready, Seedream 4.5 the most styled, but every barcode is a non-scannable fake, every nutrition panel is invented, and the best text model fakes the panel most convincingly. The box is design and AI does it well; the barcode and the panel are data and AI cannot. Composite the real functional elements, proof every word, and scan before you ship. 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 packaging from one place with the Masonry CLI.
