Here's the claim: "always upload your real photos" is bad advice, at least as a blanket rule, and following it uncritically produces worse builds than trusting the generator more often would. A 480px product shot dragged in from a five-year-old site, badly lit, off-center, is not automatically better than a clean generated stand-in just because it's "real." The instinct that authenticity beats fabrication every time is only true past a quality floor most uploaded photos don't clear. I've watched people insist on using a phone photo taken in a dim stockroom because it's "the actual product," when a generated shot in the build's actual art direction would have read better to every visitor who never met the product in person.
That's the contrarian part. Here's the mechanism that makes it true, and — since I'm not actually trying to talk you out of using your own photography — where the rule flips back the other way by the end.
Why "real" isn't the same as "better"
The image library, reachable from the + menu in the builder chat, is a flat upload surface: drag in photos, logos, product shots, screenshots, tag each with a scope — this chat only, or all sessions across your account. At build time a resolver walks the library and decides where each image lands, not by matching filenames but by reading what the image actually depicts — storefront, product on white, headshot, texture — against what each section needs. It's genuinely good at this. Upload six product photos against a six-item menu section and they land on the right cards more often than not.
But the resolver can only place what you give it, and what you give it carries real constraints that generation doesn't:
- Resolution in, resolution out — the builder never upscales, so that 480px relic ships at 480px, soft and pixelated next to everything else on the page.
- Fixed aspect ratios — a square product shot dropped into a wide banner slot gets center-cropped rather than stretched, which is correct behavior but ruthless if your subject wasn't centered to begin with.
- Accidental chrome — people upload phone-screenshot "photos," status bar and app chrome included, far more often than you'd guess; the resolver places it anyway, because it still reads as an image of the right subject, and now there's a signal bar baked into your hero.
None of that is a knock on the resolver. It's a case for being honest about your source material. A generated image, drafted in the build's own committed art direction, doesn't have any of those failure modes — no soft resolution, no off-center crop, no accidental chrome. If your real photo is bad, the "authentic" choice isn't actually the better one.
The settings that make generation coherent
Also under the + menu: two knobs that govern everything the library doesn't cover — hero backgrounds, dividers, decorative art, anything you didn't supply. First, a quality tier, and it's a real trade-off, not a default-to-max checkbox.
| Tier | Use it for |
|---|---|
| Cheap | Backgrounds and anything meant to sit behind text — iterates fast, useful when you're saying "warmer, less busy" four times running while dialing in a look. |
| Expensive | Hero-scale images, where artifacts are visible. |
My rule: draft the whole build cheap, then bump the two or three hero images to the good tier once direction's locked. Doing it backwards means paying premium cost for drafts you throw away three iterations later.
Second, style preferences — photographic versus flat, warm versus cool, busy versus minimal — which persist rather than needing restating every prompt, and which scope exactly like images do: account-wide default, chat-level override for a project that needs to diverge. An agency account shouldn't share one style preference between a law firm build and a skate shop build in the same week.
What actually produces a coherent set, though, isn't the style knob alone — it's that generated art reads from the build's committed art direction, the same one governing layout, type, and color (covered in the design director). Ask for a hero and three section illustrations in a build with a locked direction, and they share a palette and rendering style without you specifying it three times. Ask in a build with no locked direction yet, and you get plausible-but-mismatched results — a signal the direction isn't settled, not a bug.
What the "always upload real" crowd gets right
And here's the concession, because the blanket rule is wrong but the instinct underneath it isn't. There's a category of image where generation fails regardless of quality tier or art direction, and it's exactly the category people mean when they say "upload your real stuff": faces, food, and logos.
A bakery site with real pastry photos and a generated storefront exterior reads completely fine — nobody expects a generated building to be photorealistic-exact, and it's filling a genuine gap where no photo existed. The same site with a generated croissant standing in for the real one on the counter reads wrong within about two seconds, because your eye is calibrated to real food in a way it just isn't calibrated to architecture. You won't necessarily catch it yourself — you've been staring at the build for an hour — but a first-time visitor will, and that's the audience that counts.
Logos are worse, and I'll go further than "prefer real" here: a generated logo-like mark sitting in the header of an actual business is close to the worst thing the resolver can produce, and it almost always happens because nobody uploaded the real one. If you don't have a clean logo file, say so out loud rather than letting the build proceed — the alternative is a placeholder mark plausible enough that nobody catches it before launch.
So the corrected version of the rule isn't "always real" or "generation is fine, don't worry about it." It's narrower than either: upload faces, food, products, places, and logos when they exist and are decent quality — the resolver rewards good sources handsomely. Everything else, including a bad real photo, is fair game for the generator, and pretending otherwise just because a file happens to be a photograph is how you end up shipping a 480px product shot nobody would have chosen on purpose.
