A gadget with a screen has the one thing an AI image model cannot actually generate: your real interface. A smartwatch, a phone, a camera, a thermostat, the display is your software, your data, your real UI, and it is exactly the kind of dense, exact, meaningful content image models are worst at. Get the case wrong and it is an art-direction miss. Get the screen wrong and you are showing customers a product that does not exist.
So I tested it. I ran one brief, a generic smartwatch with a glossy touchscreen, a brushed-aluminum case, and a single side button, explicitly no logos or 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 cases came out great. The screens did not, and two other things showed up that matter even more. This is the electronics entry in our product-photography series, alongside the skincare, jewelry, supplements, makeup, food and beverage, footwear, candles, clothing, furniture, handbags, sunglasses, glassware, flowers, watches, perfume, 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
- For the hardware: all four produced a believable brushed-metal case, and Seedream 4.5 made the most premium macro at the lowest cost.
- For the screen: none of them. Every watch face was invented. GPT Image 2 made the most convincing fake, which makes it the most dangerous.
- The trap nobody mentions: three of four cloned the Apple Watch (case, crown, button, activity rings) despite a no-logos prompt. For a gadget, the shape is the brand.
If you only remember one thing: an AI-generated screen is not your UI, and an AI-generated gadget is often someone else's design. Use AI for the hardware, composite your real screen, and start from your real device.
The test, model by model
One brief, four models, same prompt. I judged the screen first, then the case material, then whether the model invented a recognizable brand.
GPT Image 2 is the strongest text model, and electronics is where that strength becomes a liability, exactly as it was on supplements. It produced the most detailed, most legitimate-looking screen of the four, an analog watch face with sub-dials, a step count, and a day-date, the kind of UI a customer would read as real. None of it is. The one thing it got right that the others did not: it produced a generic round watch rather than copying a famous design. Best screen realism, most dangerous fiction, cleanest on the branding question. Also the priciest and slowest.
Seedream 4.5 made the most beautiful image again, just as it did across this series, a premium macro with real brushed-titanium texture, a knurled crown, and believable glass. But it failed on both of the things that matter for electronics. The screen looks premium at a glance and dissolves into garbled near-text the moment you read it, the same way it handled small labels on skincare and supplements. And it is a near-exact Apple Watch, the rounded-square case, the crown, the side button, the activity rings, the design language Apple owns, generated from a prompt that asked for none of it. Beautiful hardware, fictional screen, borrowed identity.
Nano Banana 2 produced the most believable everyday result. The screen is a clean, legible watch face, a time, a date, a tidy grid of colorful app icons, the kind of shot you would scroll past as obviously real. That is precisely the risk: it is a convincing, fully invented interface. And like Seedream, it defaulted to the Apple Watch form, case, crown, side button, sport band, with the wordmark removed but the entire recognizable design intact. Clean photo, fake screen, cloned shape.
FLUX.2 Pro was the cheapest and gave a clean case with the sparsest screen, a few vague app icons and minimal text rather than a detailed UI. In a way that is the most honest failure: less convincing fiction. But the body is still the familiar rounded-square smartwatch silhouette, so it did not escape the trade-dress problem either. Good hardware, low-detail fake screen, same borrowed form.
The comparison
| Model | Hardware and material | Screen / UI | Avoided cloning a brand? | Rough cost/image |
|---|---|---|---|---|
| GPT Image 2 | Strong, brushed case | Most convincing, fully invented | Yes, generic round form | ~26.4 credits |
| Seedream 4.5 | Best, premium titanium | Premium look, garbled fine text | No, Apple Watch clone | ~4.8 credits |
| Nano Banana 2 | Strong, clean | Cleanest legible fake (clock + apps) | No, Apple Watch clone | ~9.3 credits |
| FLUX.2 Pro | Good | Simplest, sparse invented face | No, Apple Watch silhouette | ~3.6 credits |
Credit costs are first-hand from this test on Masonry; per-image rates move, so check current pricing.
Why electronics fail in two places at once
Most products in this series fail in one place: a label, a shade, a print. Electronics fail in two, and both are about identity rather than looks.
The screen is the supplements panel of electronics. A display is real software showing real information, and an image model has no access to either. It generates a screen-shaped block of plausible UI because that is what it saw in training, not because it is showing your product. Best case, you get an obvious placeholder. Worst case, and this is the GPT Image 2 case, you get a clean, structured, official-looking interface full of invented data that a customer reads as the real product. The prettier the fake screen, the bigger the misrepresentation.
The shape is the trademark. For a smartwatch, a pair of wireless earbuds, or a game controller, the silhouette is the brand. Apple, among others, protects that trade dress, the case proportions, the crown, the activity rings, not just the logo. A no-logos prompt strips the wordmark and leaves the recognizable design, and three of four models reached for the most famous one by default. Shipping a generated image that copies a competitor's protected design is a real risk, not a styling quirk, and it is the footwear test's invented-logo problem in a different form.
The workflow that actually works for electronics
The takeaway is not "do not use AI for gadget photos." It is "use AI for the hardware, and protect the screen and the shape."
- Let AI own the hardware and the scene. Cases, materials, lighting, lifestyle context, this is where AI saves you a studio day with its glossy-screen and metal-reflection control, and all four models did it well.
- Composite your real screen. Generate the device, then place a real screenshot onto the display rather than trusting a from-scratch UI. The screen is your software, treat it like the real asset it is.
- Start from your real device. Feed a reference photo of your actual product so the form factor is yours, not a default clone of a famous gadget. This matters more for electronics than almost any other category, because the shape carries the brand.
- Proof at full size. Zoom into the screen and the silhouette before anything ships.
With the Masonry CLI you can generate the hardware shot across models and feed your real device as a reference so the form and screen stay yours:
masonry image "smartwatch on brushed concrete, soft studio light, keep the exact device" --ref ./real-watch.png --model seedream-4-5
The bottom line
For electronics, the model choice is about the hardware, and Seedream 4.5 wins it on material and price. But a gadget is not just hardware. The screen is your real software, and every model invented it, most convincingly with the best text model. The shape is often your real brand, and most models cloned a famous one. Use AI for the device and the scene, composite your real screen, generate from your real product, and never ship a synthetic UI or a borrowed silhouette. 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 device from one place with the Masonry CLI.
