Denoising Diffusion Probabilistic Models

DDPM論文，Jonathan HoがUC Barkeleyで作った

Background
Forward process
The forward process is designed as follows:

In each step, the Gaussian noise is added. When t -> \infty, x_T converges to the Gaussian noise.

Reverse process

Loss function
Actually, the loss function is truncated because the quality of sampling is enhanced as follows:

The paper requires more time to fully understand.