It actually has impressive image generating ability, IMO. I think the two things go hand-in-hand. Its prompt adherence can be weaker than other models, though.