The cover above this article was made on my own machine. No stock site, no subscription, no server processing my request somewhere in a data center. I typed a description, waited a few minutes, and the image was on my disk. The internet did not have to be on.
That sounds unremarkable, but it breaks with how AI image generation has worked until now. And it changes who holds the control.
This is part 1 of a short series of three. I am starting with the why. In part 2 I show exactly how you do it yourself, and in part 3 I explain why I handle it the way I do.
Where image AI lives now
Almost every image generator you know runs in someone else’s data center. You type a prompt, it travels to a server, a model of tens of gigabytes sits there, and the result comes back. You pay per image or per month, and along the way your prompt and your picture pass through a company you do not control.
That works fine, until you think about it. You depend on a service that can change its price, tighten its terms, or shut down tomorrow. And the result often carries a recognizable, polished house style that everyone gets who presses the same button.
What changed
The model I use is called Bonsai Image, made by PrismML. It is based on a well-known image architecture, FLUX.2 Klein, but flattened. By reducing the weights to a little over one bit each, the model shrinks from a heavy beast to just over a gigabyte. The ternary version is 1.21 GB, the lightest one 0.93 GB.
Small enough to simply run on your laptop. On an M4 Pro it makes an image in about six seconds, several times faster than the full-size version. It even runs on an iPhone. And it is open, under an Apache 2.0 license, so you can download it and use it commercially.
The trick is not that it beats the big cloud services. Sometimes it does not. The trick is that it is good enough and fits on your own device.
Why local matters
Three things change the moment the model sits on your own machine.
Privacy. Your prompt and your image never leave your device. For a single cover that hardly matters. For work that is not ready to go out yet, or for an idea you would rather not have a service read along with, it matters completely.
Cost. There is no meter ticking per image or per month. I generate a batch, throw half of it away, try again, with no bill growing alongside. That changes how you work. You start playing instead of rationing.
Identity. Because I drive the model myself, I can hold one consistent style across everything I make. Not a generic look you see everywhere, but a visual signature of my own that fits this site.
The same idea as the rest
This fits something I write about often. A few months back a piece went into digital sovereignty, about who is really at the controls of the technology you use every day. Making images locally is a small, concrete form of that. Not a principled stand against the cloud, but the simple pleasure of keeping a part of your tooling in your own hands.
I also managed to leave WordPress behind and run my sites entirely myself. The imagery was the last piece still living in the cloud. Not anymore.
In the next part I show how to set this up yourself. It is less work than you would think.
This is part 1 of a triptych on making images locally with AI. Read on in part 2: how to make them yourself.