It does not. We have covered some JSON Experiments in other topics, but results have been inconsistent. It does indeed produce more focused and consistent outputs, and can be aided in terms of efficiency by leveraging a custom GPT, but it is more relevant for one and done generations and the same result can be produced by again just reverting the minimalist json strings to only a few words for a prompt. In terms of policy and intent, the directness of the zero-shot json structure will trip violations more often than through creative ideation.
This has been my most successful approach recently in terms of creating artistic outputs of some kind (as opposed to just testing edge cases and limits). For image gen, I’m working on a horror based graphic novel as an experiment. The fact that the model knows that, and has ideated through the context of the storyline with me, has dramatically changed the type of image outputs it will allow.
For text, I’ve been experimenting with Riffusion and using 4o for generating band concepts and lyrics. Again, with the context of the intent of the query, I am able to generate as required regardless of theme.
Off topic for this thread, but my findings have been the same. I’m down to explore further, if we want to start a new one.