Physicist's path integral paper unifies diffusion, GANs, and flow models

In a 51-page paper published on arXiv, physicist Ramon Winterhalder (affiliated with TIF-UNIMI) proposes a radical unification of generative AI models. He shows that flow-based, diffusion-based, variational, and adversarial models are not separate paradigms but different evaluation principles applied to a single master action. Using the Martin-Siggia-Rose-Janssen-de Dominicis (MSRJD) formalism—a tool borrowed from statistical field theory—he rewrites generative modeling as a path integral. This reveals a clean separation between 'free' probability flow and 'interacting' terms, which can then be treated with diagrammatic perturbation theory, the same machinery used in quantum field theory.

The practical payoff is immediate. Winterhalder derives a one-loop correction that can be applied to deterministic samplers without any additional stochastic sampling. On solvable and nonlinear drift tests, this correction reduces the tree-level error from 53% down to just 1.6%. He also treats imperfect learned scores as 'insertions' into the path integral, which naturally produces a response-weighted score-matching objective that could improve how diffusion models are trained. Furthermore, symmetry-equivariant drift design becomes an operator expansion with effective field theory (EFT) power counting, giving model architects a principled way to choose which symmetries to encode. While the paper is theoretical, its impact could be enormous: a unified language for all generative models, plus a free accuracy boost for existing deterministic samplers.