Abstract:Score-based diffusion models have proven effective in image generation and have gained widespread usage; however, the underlying factors contributing to the performance disparity between stochastic and deterministic (i.e., the probability flow ODEs) sampling schemes remain unclear. We introduce a novel formulation of diffusion models using Feynman's path integral, which is a formulation originally developed for quantum physics. We find this formulation providing comprehensive descriptions of score-based generative models, and demonstrate the derivation of backward stochastic differential equations and loss functions.The formulation accommodates an interpolating parameter connecting stochastic and deterministic sampling schemes, and we identify this parameter as a counterpart of Planck's constant in quantum physics. This analogy enables us to apply the Wentzel-Kramers-Brillouin (WKB) expansion, a well-established technique in quantum physics, for evaluating the negative log-likelihood to assess the performance disparity between stochastic and deterministic sampling schemes.

What problem does this paper attempt to address?

This paper explores a diffusion model based on Feynman path integral, which is a theory in quantum physics that aims to address the ambiguity in performance between random and deterministic sampling schemes in diffusion modeling. The authors propose a new representation method for diffusion models using the Feynman path integral formula, originally developed for quantum physics. Through this approach, they are able to provide a more comprehensive description of fraction-based generative models and derive reverse stochastic differential equations and loss functions. The paper also introduces a parameter, similar to the Planck constant in quantum physics, to connect the two sampling schemes and evaluate the differences in the negative log-likelihood of different sampling schemes using the WKB (Wentzel–Kramers–Brillouin) expansion. The main contributions of the paper include: 1. The re-expression of diffusion models using path integral techniques, which deepens the understanding of the mathematical properties of these models and allows for the application of various techniques from quantum physics. 2. The introduction of an interpolation parameter h that connects random generation (h=1) and probabilistic flow ODE (h=0). The path integral framework reveals the role of h in diffusion models similar to the Planck constant in quantum physics. 3. The application of the WKB expansion method to quantitatively analyze the benefits of noise in the sampling process by conducting a first-order expansion of the negative log-likelihood calculation with respect to h. The paper also discusses the differences between random and deterministic sampling processes and the role of noise in the sampling process, but lacks in-depth quantitative analysis or a rigorous theoretical framework to explain this phenomenon. By drawing analogies with quantum physics, these contributions reveal a deep connection between diffusion models and physics, going beyond the perspectives of classical Brownian motion.

Understanding Diffusion Models by Feynman's Path Integral

Score-based Diffusion Models via Stochastic Differential Equations -- a Technical Tutorial

Stochastic quantization and diffusion models

Generative Fractional Diffusion Models

Random Walks with Tweedie: A Unified Framework for Diffusion Models

How Discrete and Continuous Diffusion Meet: Comprehensive Analysis of Discrete Diffusion Models via a Stochastic Integral Framework

A Flexible Diffusion Model

Lecture Notes in Probabilistic Diffusion Models

Fourier Diffusion Models: A Method to Control MTF and NPS in Score-Based Stochastic Image Generation

Stochastic Interpolants: A Unifying Framework for Flows and Diffusions

Nonequilbrium physics of generative diffusion models

Closing the ODE-SDE gap in score-based diffusion models through the Fokker-Planck equation

A Sharp Convergence Theory for The Probability Flow ODEs of Diffusion Models

The Unreasonable Effectiveness of Gaussian Score Approximation for Diffusion Models and its Applications

An Overview of Diffusion Models: Applications, Guided Generation, Statistical Rates and Optimization

On the Generalization Properties of Diffusion Models

Diffusion models for Gaussian distributions: Exact solutions and Wasserstein errors

Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models

Analyzing Neural Network-Based Generative Diffusion Models through Convex Optimization

Towards Faster Non-Asymptotic Convergence for Diffusion-Based Generative Models

FreeFlow: A Comprehensive Understanding on Diffusion Probabilistic Models via Optimal Transport