Generating Images using VAEs, DCGANs, and DDPMs (Diffusion)

Overview

In this project, we will explore 3 image generation models: VAE, DCGAN, and DDPMs (diffusion). For each model, we discuss the theory and intuition before implementing each in Keras. Then we will generate images in the style of MNIST handwritten digits, which will allow us to compare and contrast the different models using the Generative Model Trilemma. Check out our blog post!

Reference

The implementation of the three models are heavily based on the followig sources!

VAEs

1. https://www.tensorflow.org/tutorials/generative/cvae
2. https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73
3. https://arxiv.org/pdf/1907.08956.pdf

DCGANs

1. https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html
2. https://towardsdatascience.com/dcgans-deep-convolutional-generative-adversarial-networks-c7f392c2c8f8

DDPMs

1. https://learnopencv.com/image-generation-using-diffusion-models/
2. https://keras.io/examples/generative/ddpm/
3. https://medium.com/@vedantjumle/image-generation-with-diffusion-models-using-keras-and-tensorflow-9f60aae72ac