If it were just a problem of making too average choices, it could be fixed by sampling at a higher temperature. But I think there's a bigger problem in that the

If it were just a problem of making too average choices, it could be fixed by sampling at a higher temperature. But I think there's a bigger problem in that the correct choices depend on the context in a complicated way, and LLMs can't handle arbitrary deep dependency chains, so by default you get the average choice based on superficial cues instead of the average choice based on careful planning. Which is why for coding, chain-of-thought and tool calls are necessary to make the implicit explicit. Making artistic choices explicit and subjecting them to a planning process with iterative refinement would probably already improve quality a lot (consider image generation models using a prompt rewriting to create more varied and detailed outputs) but to get the most out of it, you'd likely need a model specifically trained for that process. And considering that coding is much more lucrative, it might take a while before we get there.