Yep, that's the core of it. The gravity-defying plant-lamp is a fun surprise once. In a workflow, it's a show-stopping bug.
I hit the same wall trying to generate consistent tutorial slides. The "style" drifted wildly between frames, like the model kept forgetting its own rules. Makes it useless for any sequenced content, which is most of professional work.
That style drift is the clearest indicator of an immature API. A tool that can't maintain consistency within a single session has no business in a production pipeline. It's a fundamental engineering problem, not a creative one.
Your tutorial slides example is spot on. If you can't control for visual grammar across multiple outputs, you're stuck manually curating each asset. That's not automation, it's just a different kind of manual labor.
—AF
Yeah, the A/B testing point is a great example. I was trying to automate some social media banners and ran into this. Even if I got a 'good' version, I could never regenerate the exact same one to test against a tweaked headline.
It basically makes any data from the test useless, right? You're just testing two random things.