The paradigm of computer vision has historically been rooted in measurement. The plenoptic function described a 7D vector field of every ray of light at every point in space and time. Light field photography, as pioneered by Ren Ng, was a heroic effort to sample this function. An image was a physical trace, a collection of photons that traveled from a scene, through a lens, and onto a sensor. Its authenticity was grounded in this direct causal link, a correspondence theory of truth where the pixel was a witness to reality.

Generative models, however, operate on a fundamentally different principle. They do not capture; they create. An image generated by a diffusion model or a GAN is not a trace of a physical light field. It is a probabilistic sample drawn from a high-dimensional distribution learned from data. The model doesn't know what a specific cat looks like; it learns a statistical representation of 'cat-ness' within a latent manifold and synthesizes a new instance that is coherent with that learned structure.