Generation or Editing?
I Was Asking the Wrong Question.
Every time I opened an AI image tool, I lost a few seconds to the same small decision: am I generating or editing? It sounds trivial. It wasn't. I went in circles on this for longer than I'd like to admit. Then it clicked — and the answer is almost annoyingly simple.
It's all generation
At the level that matters — what the model is actually doing — there is only one thing happening: the model generates pixels from whatever you give it. That is the whole engine. It never does anything else.
"Generation" and "editing" are not two different machines. They are two workflows running on the same machine. The word "edit" describes what you are doing — changing something that already exists. It does not describe what the model is doing, which is always the same: generating.
Once I saw that, the confusion dissolved. Brushing over a region in an image isn't a different mode. It's just one more instruction — you're telling the model where it's allowed to generate. Still generation. Just pointed at a smaller area.
What actually changes: the dial
If the engine is always generation, then the real difference between "generating" and "editing" is something else entirely: how much of an existing image you feed back in as a constraint.
That's not a switch with two positions. It's a dial.
Same engine, the whole way down. The only thing moving is how tightly you're pinning it to something that already exists. Low constraint — we call it generation. High constraint — we call it editing. Everything useful happens on the dial in between.
The question that replaced the old one
So I stopped asking "is this generation or editing?" — because the honest answer is always "both, it's generation." That question was never going to resolve, which is exactly why I kept getting stuck on it.
The question that actually pays off is: how much of an existing image do I need to keep?
Nothing to keep? You're at the free end. Describe it and let it run. A real product, a real face, a real layout that has to survive untouched? You're at the constrained end. Give the model the image and pin down everything you're not changing.
That single question — how much do I need to keep? — routes almost every decision I used to fumble. You are no longer choosing a mode. You are choosing a position on the dial.
One catch — the tool still matters
Here's the part that keeps this from being a neat "it's all the same, relax" story.
Mechanically, yes — it's all generation. But models are trained for different ends of the dial. A pure text-to-image model, pushed to edit, will quietly redraw things you wanted left alone — because it was built to invent, not to preserve. A model trained specifically for editing is taught to hold the rest of the image still while it changes the one thing you asked for.
The short version
The industry sells these as two features with two buttons. Under the hood it is one engine and a constraint dial. The labels "generate" and "edit" describe your workflow, not the model's mechanism.
Once you stop asking "which mode is this?" and start asking "how much do I need to preserve?", the guessing stops — and you start picking the right tool on purpose instead of by trial and error.
That reframe cost me a few weeks of low-grade confusion. Hopefully it costs you about five minutes.
Published by IMA AI — June 2026.