A supplement bottle has the one thing AI image models cannot legally fake: a Supplement Facts panel. It is FDA-regulated under 21 CFR 101.36, with required content, a boxed format, a set order, and minimum type sizes. It is, in other words, exactly the kind of dense, exact, regulated text that image models are worst at, and the kind of text where getting it wrong is not an art-direction miss but a compliance problem.
So I tested it. I ran one brief, an amber glass supplement bottle labeled "VITALITY" with a Supplement Facts panel and a few capsules, through the strongest image models with the same prompt. The photos came out great. The labels did not. Every model invented the Supplement Facts panel, and the result that should worry you most came from the model that is best at text. This is the supplements entry in our product-photography series, alongside the skincare and jewelry tests and the broader best AI image model for product photography roundup.
Quick answer
- For the photo: all three models I got results from produced a believable bottle, and Seedream 4.5 made the most premium shot at the lowest cost.
- For the label: none of them. Every Supplement Facts panel was fabricated. Use AI for the scene and keep your real label.
- The trap: GPT Image 2 made the most convincing-looking panel, which makes it the most dangerous, because it is entirely invented and passes a glance.
If you only remember one thing: an AI-generated Supplement Facts panel is not your formula. It is a plausible fake. Never ship one.
The test, model by model
One brief, same prompt. I judged the photo first, then zoomed into the label, which is where supplements are won or lost. (FLUX.2 Pro errored on this brief across two attempts and is not shown.)
GPT Image 2 is the strongest text model, and supplements is where that strength becomes a liability. It produced a Supplement Facts panel that looks legitimate, boxed, structured, with an ingredient table and a percent-daily-value column. None of it is real. A panel that reads as compliant but is fabricated is worse than obvious gibberish, because it is the one a busy team might actually ship. If you take nothing else from this test: the better a model is at text, the more carefully you have to stop it from writing your label.
Nano Banana 2 gave the most useful split for understanding the problem. The front-of-bottle label, brand name, product line, capsule count, came out clean and legible, the same simple-label strength it showed on skincare. But the Supplement Facts panel was a boxed placeholder filled with unreadable text. In a way this is the safer failure: nobody would mistake it for a real panel. The photo and the front label are usable; the regulated panel is not.
Seedream 4.5 made the best image again, just as it did on skincare and jewelry, a genuinely premium macro shot, real amber glass with the capsules visible inside, at the lowest cost of the three. The brand name is clean. The Supplement Facts panel, wrapping around the side, is garbled near-text with invented ingredient lines. Beautiful bottle, fictional label.
The comparison
| Model | Photo quality | Front label | Supplement Facts panel | Rough cost/image |
|---|---|---|---|---|
| Seedream 4.5 | Best, premium macro | Clean brand | Garbled, invented | ~4.8 credits |
| Nano Banana 2 | Good | Clean and legible | Boxed but gibberish (obvious fake) | ~9.3 credits |
| GPT Image 2 | Good, bright studio | Clean | Convincing but fully invented (risk) | ~26.4 credits |
| FLUX.2 Pro | Errored on this brief, not shown | - | - | ~3.6 credits |
Credit costs are first-hand from this test on Masonry; rates move, so check current pricing.
Why supplements are a compliance problem, not just a photo problem
For most products, a wrong label is an art-direction miss. For supplements it is a regulatory one. The Supplement Facts panel is one of the most strictly governed elements of any consumer package. The FDA's dietary supplement labeling rules require a specific set of elements, the identity, net quantity, the Supplement Facts panel itself, the ingredient list, the business name and address, allergen declarations, and more, and the panel has to be boxed, in a single legible type, at least 1/16-inch, with a set nutrient order and no filler breaking up the required items. Structure-function claims drag in a mandatory disclaimer and an FDA notification.
An image model knows none of this. It produces a panel-shaped block of plausible text because that is what it saw in training, not because it is reporting your formula. Best case, you get obvious gibberish you would never ship. Worst case, and this is the GPT Image 2 case, you get a clean, structured, official-looking panel full of ingredients your product does not contain and amounts you never tested. Shipping that is misbranding. The prettier the fake, the bigger the trap.
The workflow that actually works for supplements
The takeaway is not "do not use AI for supplement photos." It is "use AI for the part it is good at, and protect the part it is not."
- Let AI own the scene. Backgrounds, lighting, lifestyle context, seasonal concepts, the composition, this is where AI saves you a studio day, and all three models did it well.
- Keep your real label. Start from a real photo of your bottle and have AI replace or restyle the background, or feed your actual label as a reference image, so the panel in the final shot is your true, compliant one. This is exactly why most supplement-specific tools are background-replacers rather than from-scratch generators.
- Never generate the panel. No model, not even the best text model, should write your Supplement Facts panel. Treat it like the regulated document it is.
- Proof at full size. Zoom into every label before it ships, the same discipline as any AI output, with higher stakes.
With the Masonry CLI you can run the same scene prompt across models to pick the best photo, while feeding your real product image so the label stays yours:
masonry image "amber supplement bottle on marble, soft studio light, lifestyle background" --image ./real-bottle.png --model seedream-4-5
The bottom line
For supplements, the model choice is about the photo, and Seedream 4.5 wins it on look and price. But the label is not a thing you generate, it is a thing you protect. Every model in this test invented the Supplement Facts panel, and the best text model invented it most convincingly. Use AI for the scene, keep your real label, and never ship a synthetic panel. See how the same fidelity-first logic plays out across products in our best AI image model for product photography roundup, or run your own bottle from one place with the Masonry CLI.
