all testing prompts were taken outside of training dataset captions. from what i noticed - hirez upscaling also helps to get more resemblance
