You're describing the exact moment you realize the tool is an asset generator, not a scene generator.
> Has anyone else run into this?
We've all run into it. Your last line is the key. You can't integrate it for *that* asset type. The only prompt that works is "generate background setting, no people."
The safe data thing is a red herring. Other models with less restricted data have the same fundamental issue with articulated forms. They just make different weird mistakes.
Your handling method is already in your post: you have a usual workflow with other tools. Use that for the human component. Firefly becomes your background plate generator. Trying to force it to do both is where the hours vanish.
—cp
Yeah, that "fantastic coherence for objects and scenes" is the trap. It's like getting a perfect, fully-furnished office and then discovering the chair is a geometry-defying blob.
To your questions: no, I haven't found a prompt fix for anatomy. It's not a style modifier problem, it's a physics problem the model doesn't have. The safe data line is a smokescreen. Even models trained on everything make weird hands, they just make *different* weird hands.
My handling method is the same as the cost calculation: use it for the cafe background, then buy the cheapest stock human that fits. The composite step is a fixed, predictable cost. Trying to fix a malformed joint in post always ends up being a time-and-money sink that blows any budget you had for "quick asset creation."
Cloud costs are not destiny.