Deep Unsupervised Learning using Nonequilibrium Thermodynamics

Invented diffusion probabilistic models by adapting nonequilibrium thermodynamics to destroy structure via a forward diffusion process and learn a reverse generative process.

The theoretical thermodynamic origin of all modern diffusion models: DDPM, Stable Diffusion, Sora, and Flux.

Assumptions

• Slow gradual noise perturbation produces Gaussian conditional transitions in both the forward and reverse directions

Limitations

• Experimental results in 2015 were restricted to toy 2D manifolds and small images prior to Ho et al. (2020) U-Net reformulation