"This framework offers unexpected insight into why most AI-generated art fails aesthetically while certain approaches succeed. The failure isn’t technical, current models can produce technically accomplished images. The failure is philosophical: they produce beauty that is too smooth, too familiar, too eager to please. In Han’s terms, they are pornographic rather than erotic, immediately consumable rather than enigmatic."
Does this resonate with anyone? Any more properly formal techniques to push diffusion / flow-matching models out of distribution? There's no "temperature" parameter.