A Denoising Diffusion Probabilistic Model (DDPM) is a type of deep generative latent model which generates new data by iteratively adding and removing noise from some prior distribution data [1]. A DDPM generally:

1. Gradually corrupts the original data with noise in a controlled way during the forward process
2. Allows the model to learn to denoise effectively during training
3. Enables high-quality sample generation during the reverse process

The implemented DDPM uses a simple CNN based noise predictor model which learns to predict and remove noise from images. It takes a noisy image and timestep as input and predicts the noise that was added, enabling the gradual denoising process.

Some of the best generated images are shown below:

Code Implemetation

References

[1] Ho, J., Jain, A., & Abbeel, P. (2020). Denoising Diffusion Probabilistic Models. https://arxiv.org/abs/2006.11239