14 September 2026
Hi, I'm Barbara. Want to know why this blog exists?
A day job, and the AI at night. Who I am, why I opened this site, and what you'll find every time you come back: my method, work in progress.
Blog · Method
For months I blamed the tool when an image came out fake. Then I understood: the AI isn't the one getting it wrong, I am, in how I ask. Here's what I changed.
For months I made the same mistake, and I think a lot of people do. When an image came out fake — plasticky, shop-window lighting, that indefinable "made-on-a-computer" look — I blamed the tool. I switched models, tried a new one, pinned my hopes on the next version. Then I ran a test that took away my excuse: I kept the same tool and the same scene, and changed only one thing — what I was asking for. The photo became a different photo.
I realised that the AI, almost always, doesn't get the image wrong: it executes a badly made request badly. "Generate a photo of the product" isn't a brief, it's a wish. The model fills the gaps however it likes, and there are a lot of gaps: where does the light come from? What is the object made of? What's in the background, and what shadows does it cast? If you don't tell it, it makes it up — and usually in the most generic way possible.
Now, before I press "generate," I write three things I used to take for granted. The first is the light: a single source, where it comes from, soft or hard. The second is the materials: matte, satin, ceramic, glass — because that's where realism is won or lost, in the way a surface catches the light. The third is the background and consistency: the same object, the same details, the same scene from one image to the next, so it doesn't change its face every time.
With Nano Banana, the same product photo that used to look like a fake render became a catalogue-grade shot, simply because I stopped letting the machine decide.
The more precise I am in the brief, the more real the result looks. It sounds obvious, but it's the opposite of how we're tempted to use these tools: fast, at random, hoping. The time you save writing a lazy prompt you pay back in twenty regenerations.
Since I understood this, I've stopped chasing the perfect tool. What I already have in my hands is almost always enough. I was the one not telling it what I saw in my head.
Read next
14 September 2026
A day job, and the AI at night. Who I am, why I opened this site, and what you'll find every time you come back: my method, work in progress.
18 September 2026
The video model I use myself in Hollywood's crosshairs: suspended launch, watermarks, blocks on real faces. We're at 2.5 now, and that deal covers it too.