Toys are the category where AI's greatest strength, its memory of what things look like, becomes a legal liability. Ask for almost any toy and the model does not invent a generic one, it reaches for the most famous version it has seen. And the most famous versions of toys are among the most fiercely protected designs in the world.
So I tested the cleanest possible version of this. I ran one brief, generic colorful plastic interlocking building bricks, explicitly no logos and no branding, 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 plastic and color came out great on all four. And all four returned LEGO's unmistakable stud-and-tube trade dress, the exact protected design I had asked them not to brand. This is the toy 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, packaging, pet products, 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
- The finding that matters: all four models cloned LEGO's stud-and-tube design from a no-branding prompt. For toys, the model reaches for protected trade dress by default.
- On craft: glossy plastic and accurate colors on all four, with Seedream 4.5 the most premium at the lowest cost. But craft is not the decision here.
- The rule: never sell an AI toy image generated from a text prompt. Start from your own design as a reference, and get legal review for commercial use.
If you only remember one thing: for toys, a no-branding prompt does not protect you. The shape the model picks is usually someone else's trademark, so the toy has to be your design, fed as a reference, not invented by the model.
The test, model by model
One brief, four models, same prompt. The craft question is quickly answered, all four are good, so the real comparison is how unmistakably each cloned the protected design.
Seedream 4.5 made the most beautiful bricks: tactile glossy plastic with realistic highlights and vivid red, blue, yellow, and green. It is also the clearest illustration of the problem, because the better the render, the more obviously it is LEGO. The stud-and-tube system, the proportions, the way the bricks snap together, this is LEGO's protected trade dress rendered in premium detail. Best craft, and a design you cannot sell.
Nano Banana 2 produced a clean, catalog-ready pile of bricks with accurate colors and good plastic. And like the others, the bricks are pure LEGO: the classic proportions, the studs, the interlocking shapes. It is a perfectly nice image of a design that belongs to someone else, which is exactly the trap.
GPT Image 2 rendered bright, glossy bricks, with some studs even carrying molded detail of the kind real branded bricks have. The craft is strong and the colors are accurate. It is also unmistakably the LEGO system, and the molded stud detail makes the trade-dress resemblance, if anything, sharper. The priciest model, and no safer on the issue that matters.
FLUX.2 Pro was the only one that drifted at all, with marginally rounder, softer studs that read as a touch more generic. But it is still plainly the LEGO stud-and-tube system, just slightly less crisp. At the lowest cost it is a fine render, and on the trade-dress question it is barely a step away from the rest. None of the four gave a genuinely original brick.
The comparison
| Model | Plastic + color | Design produced | Sellable as-is? | Rough cost/image |
|---|---|---|---|---|
| Seedream 4.5 | Best, tactile gloss | LEGO stud-and-tube clone | No | ~4.8 credits |
| Nano Banana 2 | Strong, accurate | LEGO clone | No | ~9.3 credits |
| GPT Image 2 | Strong, molded detail | LEGO clone (sharpest) | No | ~26.4 credits |
| FLUX.2 Pro | Good, softer | LEGO-style, marginally generic | No, still risky | ~3.6 credits |
Credit costs are first-hand from this test on Masonry; per-image rates move, so check current pricing.
Why toys are a trademark problem, not a craft problem
Every other category in this series fails on a detail you can fix. Toys fail on the whole object, and the fix is not a better prompt.
The model reaches for the famous design by default. Iconic toys, LEGO bricks, Barbie, Funko Pop, vintage action figures, dominate the training data, so when you ask for a category, the model returns the archetype, and the archetype is trademarked. This is the same trade-dress reflex that produced an Apple Watch, a TAG-Carrera watch, and an Adidas sneaker elsewhere in this series, but for toys it is sharper, because the entire product is the protected design, not a logo on it.
This is a documented, real risk. It is not theoretical: LEGO itself had to apologize after AI-generated artwork for one of its lines inadvertently included unlicensed third-party intellectual property. If the company that owns the bricks can ship accidental IP through AI, a small seller prompting for "building bricks" is squarely in the same trap. Guidance for the popular AI-action-figure trend says it plainly: do not generate figures that imitate trademarked designs for commercial sale.
A no-branding prompt does not save you. I asked for no logos and no branding, and got LEGO anyway, because trade dress is the shape, not the wordmark, exactly as it was for the smartwatch. Removing the logo from the prompt does not remove the protected design from the output.
How to use AI for your toy line safely
The workflow for toys is different from every other category, because the risk is legal, not aesthetic.
- Start from your own design. Generate or photograph your actual toy, then feed it as a reference so AI stages and lights your design rather than inventing a famous one.
- Avoid archetype prompts. Do not name or imply iconic formats, no "LEGO-style," "Barbie box," or "Funko." They pull the model straight to protected designs.
- Check every result. Compare against known trademarked toys before using anything commercially, and get legal review for products you sell.
- Use AI for the scene, not the toy. Backgrounds, lighting, lifestyle, packaging mockups around your real product, that is where the savings are safe.
With the Masonry CLI you can stage your own toy from a reference rather than letting a prompt invent one:
masonry image "place this exact toy on a bright playroom rug, soft daylight, photoreal" --ref ./my-toy.png --model seedream-4-5
The bottom line
Toys break the pattern of this series. Everywhere else the question is which model renders your product best; for toys the question is whether the model will hand you a product you are allowed to sell, and from a text prompt the answer was no on all four. The plastic and color are excellent, Seedream 4.5 best, but every result was LEGO's protected design, which is the only fact that matters. Start from your own toy as a reference, avoid archetype prompts, and get legal review before anything commercial. See how the same fidelity-first logic plays out across every product type in our best AI image model for product photography roundup, or stage your own toys from one place with the Masonry CLI.
