Diffusion Augmented Flows: Combining Normalizing Flows and Diffusion Models for Accurate Latent Space Mapping

We present a novel generative model that combines the strengths of two well-known generative models: normalizing flows and diffusion models. Normalizing flows lack the ability to fully map to Gaussian space, resulting in limited expressiveness. To overcome this, our model utilizes normalizing flows to map the complex data distribution to an intermediate latent distribution and then employs a diffusion model to further push forward it to the latent distribution. In addition, we introduce a new training procedure that combines maximum likelihood estimation from normalizing flows and variational lower bound from diffusion models, resulting in a unified end-to-end architecture. We evaluate our model using the Fréchet Inception Distance and Negative Log-likelihood scores and show that our model outperforms Neural Spline Flows and gives comparable results to traditional diffusion models. Our work presents a promising direction in the field of generative modeling, specifically in image synthesis.