> it has been trained to produce high-quality outputs, not to ensure the validity of mappings.
Just nitpicking here a bit. It has been trained to ensure the validity of mappings, but only for mappings of valid prompts, where "valid" is vaguely "things that appear in the training set". On the other hand, it wasn't trained to ensure the validity of mappings of invalid prompts to images. It's an open question what that would even mean - someone in another thread here suggested it should output "not a valid prompt" in this case.
Just nitpicking here a bit. It has been trained to ensure the validity of mappings, but only for mappings of valid prompts, where "valid" is vaguely "things that appear in the training set". On the other hand, it wasn't trained to ensure the validity of mappings of invalid prompts to images. It's an open question what that would even mean - someone in another thread here suggested it should output "not a valid prompt" in this case.