Footwear is a deceptively hard product to generate, and the hard part is not what most people expect. A sneaker mixes several materials in one object, knit-mesh, suede, leather, rubber, and it carries branding that is legally yours or legally someone else's. So I ran one test that checks both at once: I asked four image models for a generic running sneaker, white mesh, tan suede, white rubber sole, explicitly "no logos or branding," and watched whether they could render the materials and whether they would obey the rule.
The materials were easy. The branding was not, and that is the finding that matters. This is the footwear entry in our product-photography series, alongside the skincare, jewelry, supplements, makeup, food and beverage, candles, clothing, and furniture tests and the broader best AI image model for product photography roundup.
Quick answer
- Materials are not the problem. All four models rendered distinct mesh, suede, and rubber convincingly.
- Branding is the problem. Three models obeyed "no logos." One, Seedream 4.5, invented a real brand's trademark, stripes and wordmark, despite the instruction.
- Best material: FLUX.2 Pro, and it stayed brand-free. Best all-rounders: Nano Banana 2 and GPT Image 2, both clean and logo-free.
- The rule: keep prompts brand-free, inspect every shoe for invented marks, and never ship a hallucinated logo.
The test, model by model
Same brief, four models: a generic running sneaker, mesh and suede and rubber, no logos. Judged on the material mix and, more importantly, on whether any branding appeared.
FLUX.2 Pro rendered the most convincing materials, the mesh reads as mesh, the suede has nap, the rubber sole looks like rubber, and it kept the shoe completely generic. This is the same material-fidelity strength it showed on the drink can, and here it pairs with the safest output: no invented branding.
Nano Banana 2 gave a clean, accurate, brand-free sneaker with a chunky sole that matched the brief. The dependable middle it has been across this whole series, and it obeyed the no-logos rule without fuss.
GPT Image 2 produced a clean, generic sneaker with distinct materials and no branding, in a slightly warmer grade. Solid and safe, just the most expensive option and not noticeably better than the cheaper models here.
Seedream 4.5 made the most photographic image of the four, and it is exactly why you cannot judge footwear on looks alone. The model ignored "no logos" and rendered a real, recognizable brand's three stripes and wordmark onto the shoe. The craft is excellent. It is also a trademark you do not own, generated onto a product you were trying to keep generic. Beautiful, and the one output here you absolutely cannot publish.
The comparison
| Model | Material mix | Obeyed "no logos" | Hero look | Rough cost/image |
|---|---|---|---|---|
| FLUX.2 Pro | Best, photoreal | Yes | Editorial | ~3.6 credits |
| Nano Banana 2 | Strong | Yes | Balanced | ~9.3 credits |
| GPT Image 2 | Good | Yes | Clean, warm grade | ~26.4 credits |
| Seedream 4.5 | Strong | No, invented a real brand | Best, but unusable | ~4.8 credits |
Credit costs are first-hand from this test on Masonry; rates move, so check current pricing.
The real lesson: the logo is the landmine
Across this series, each category has had one thing the model must not fabricate. For supplements it was the regulated Facts panel. For makeup it was the exact shade. For footwear it is the branding, and it is arguably the most dangerous because it can look perfect.
A model trained on millions of product photos has seen a lot of branded shoes, so when you ask for "a sneaker," its instinct is to make one that looks like the sneakers it knows, complete with the marks. "No logos" helps, but as this test shows, it is not a guarantee. A hallucinated real-world trademark on your product image is not a quality bug, it is a legal exposure, and the prettier the render, the more likely it slips through review.
So the workflow for footwear is specific:
- Keep prompts brand-free and say "no logos," but treat that as a request, not a guarantee.
- Inspect every output for invented marks, stripes, swooshes, wordmarks, heel tabs, especially on the most beautiful results, which are the most likely to have confidently added them.
- When the branding has to be yours, use your real shoe as a reference image so the model preserves your marks instead of inventing someone else's.
- For materials and angles, you are in good shape. The mesh, suede, and rubber all rendered well, so AI is genuinely useful for multiplying angles and colorways once the branding is controlled.
With the Masonry CLI you can run the same shoe across models and compare both the material and the branding behavior in one pass, feeding your real shoe so the marks stay yours:
masonry image "three-quarter studio shot of a running sneaker on concrete, soft light" --image ./real-sneaker.png --model flux-2-pro
The bottom line
For footwear, every model could render the materials, and that is no longer where the risk is. The risk is branding: the best-looking model in this test invented a real trademark on a shoe that was supposed to be generic. Use AI for footwear, lean on FLUX.2 Pro for material realism, but inspect every output for invented marks and keep your real branding in the loop with a reference. Run yours from one place with the Masonry CLI, or see the full series in our best AI image model for product photography roundup.
