AI models don't simulate optics. They pattern-match visual descriptions. Here's what actually triggers wide-angle distortion, telephoto compression, and editorial perspective — tested prompt by prompt.