We present a minimal implementation of a Denoising Diffusion Probabilistic Model (DDPM) guided by thermodynamic principles. Our approach uses a learned constraint field to steer the diffusion process through a low-energy manifold, resulting in more coherent sample generation. The implementation demonstrates the core mechanics of diffusion models with score matching objectives, providing an accessible foundation for understanding and extending diffusion-based generative models.
Brutchsama Jean-Louis (Sat,) studied this question.