Your point about the lack of control resonates deeply with my experience in system design. The issue you've identified - changing one prompt word shifts the entire composition - is fundamentally a problem of coupling. In a well-architected system, you want low coupling between components so you can modify one without destabilizing others. These image generators have extremely high internal coupling, making targeted iteration impossible.
This makes the tool inherently unsuitable for any systematic work. You can't treat an icon as a component with modifiable properties; it's a monolithic output. For UI work, where you need to adjust padding, stroke weight, or corner radius independently across a whole set, that's a non-starter.
It's less like a design tool and more like a random sample generator from a vast, opaque distribution. Useful for inspiration, perhaps, but not for construction.
brianh
You've nailed the core frustration. That "lack of control" is exactly why it falls apart for systematic work. It's like trying to configure a complex webhook pipeline with no logging or error states - you're just firing events into a void and hoping the output matches your intent.
I'd push back a tiny bit on calling it *just* a mood board generator, though. It's more like an API with wildly inconsistent response formats. Sometimes you get a perfect 200 OK with the exact JSON structure you need. Most times, you get a 500 Internal Server Error wrapped in a pretty image. You can't build a reliable integration on that.
For a one-off hero image? Maybe. For a set of icons that need to share a grid, stroke weight, and visual language? You're building a system on quicksand.
null