Date

2022-8-31

Author

Sayın, Deniz

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

361
views

252
downloads

Cite This

Generative models have been shown to be able to produce very high fidelity samples in natural image generation tasks in recent years, especially using generative adverserial network and denoising diffusion model based approaches. Normalizing flow models are another class of generative models, which are based on learning invertible mappings between the latent space and the image space. Normalizing flow models possess desirable features such as the ability to perform exact density estimation and simple maximum likelihood based training, which can offer theoretical guarantees. While the state-of-the-art normalizing flow models are able to produce high fidelity images on specific simple image generation tasks such as faces and bedrooms, they typically fail to produce sensible results in difficult natural image datasets containing a multitude of underlying classes. We propose an approach focused on improving natural image generation using a new normalizing flow model, in which we start by generating a small natural image and refine it step by step with conditional normalizing flow models performing 2x super-resolution. We also propose a new augmentation method at the feature level for conditional encodings to make the intermediate models in our cascade more robust against noise and artifacts coming previous levels of the cascade. This augmentation method has its roots in variational inference. We perform experiments on the CelebA and CIFAR-10 datasets, show our qualitative results and compare our generations with state-of-the-art approaches using the FID metric.

Subject Keywords

generative models, natural image generation, normalizing flows, variational inference

URI

https://hdl.handle.net/11511/99466

Collections

Graduate School of Natural and Applied Sciences, Thesis