“The key differentiator for state-of-the-art image models is now the quality of their worst outputs ('lemons') for a given prompt, as most models can produce high-quality 'cherry-picked' examples.”