I'm also having issues. Every day I get model overloaded errors and it hasn't improved at all since they released NBP. Today it's almost completely broken with >80% generations failing with model overloaded errors.
Very cool! I think you're on to something here but I don't feel it's interesting/good enough yet to move me away from OpenAI's shitty chat interface.
I do some prompt engineering on a daily basis, and I find that the challenge is often less about the writing and more about ensuring that the output is what I want, and consistently so.
Trying to think of some useful features for this tool, I'm thinking that some improved tooling for testing output could be useful. One thing is to validate the output, but also to be able to run the same query in bulk to check that the output always validates correctly. The model often outputs something correctly one time, and incorrectly another time, for the same prompt.
I'm not sure I agree that the "after" examples are better than the "before" examples on your website - they're just different styles.
I tried to run it with the following query:
Hyper-realistic, aluminum panels, shiny windows, blue sky with clouds, lush nature, brick driveway to garage, grass garden, morning light, modern design, architecture design, environment, Cinematography, mega scans, cinematic, hyper-realistic, photo real, cinematic composition, highly detailed, vray,8k render, 16 mm lens, DSLR
And the only thing it changed, was by adding "--ar 3:2". Not really that much of an improvement over the original prompt. The above prompt is maybe not that bad already but for this tool to be useful, I think it should be able to improve even the good ones to be even better.