Building the state-of-the-art generative model from scratch
The code deploys the key aspects of the 2020 research publication on DDPM by Jonathan Ho (https://arxiv.org/abs/2006.11239) along with the improved technique of classifier-free guidance, as suggested by Ho and Saliman's 2021 publication(https://arxiv.org/abs/2207.12598) to sample images with higher fidelity than the ones generated by vanilla DDPM.
Reference to a detailed statistical analogy of the design and methodology:
https://lilianweng.github.io/posts/2021-07-11-diffusion-models/#classifier-free-guidance