Diffusion Models and Path Integrals

In recent years, AI-based image generation has attracted much attention. Behind AI producing photorealistic images and artistic illustrations lies a technique called the "diffusion model": images are gradually blurred into noise, and a sharp image is generated by retracing the reverse process. This seminar first explains the basic operating principle of diffusion models. Then, as recent work of ours, we introduce an approach that reformulates and analyzes diffusion models from a new perspective, using the path-integral formulation of quantum mechanics.