Denoising Diffusion

This repo implements the Denoising Diffusion Probabilistic Model(DDPM) on the MNIST dataset. The repo is just for me to better understand the DDPM so do not expect a high-performance model and do not trust the results too much. For someone who wants to re-implement the DDPM that is exactly the same as the model proposed in the the paper, I would recommend to check out the code from diffusers. In my repo, I only implement a vanilla Unet for modeling the $\epsilon_{\theta}$ without too much fascinating tricks. With such a simple model, I suppose the model can only deal with images with low resolution such as MNIST. I recently tried to add more features to the model to enhance the representation power of the model e.g. using the PixcelCNN and the idea from Stable Diffusion. So, there is a lot of uncompleted work in this repo. I will try to update the repo in the future.

Quick Start

Train

python train_ddpm.py --config config/ddpm_mnist.yaml --task mnist

Inference The anime gif and the image

python sample.py

Results

Images shown below is generated from a model which is trained on MNIST for 10K steps. Note that the result is not good enough. More iterations are needed to get a better result.

Generated images

Loss curve

Reference

[1] Denoising Diffusion Probabilistic Models