Prompt adherence and photorealism do not always improve in the same direction; checkpoints fine-tuned on large volumes of real photographs excel in skin texture and lighting, but may have weaker comprehension of complex prompts than base models.