Fast & Efficient Normalizing Flows and Applications of Image Generative Models

This thesis presents novel contributions in two primary areas: advancing the efficiency of generative models, particularly normalizing flows, and applying generative models to solve real-world computer vision challenges. The first part introduces significant improvements to normalizing flow architectures through invertible 3x3 convolutions, a more efficient Quad-coupling layer, a fast parallel inversion algorithm for kxk convolutional layers, a fast backpropagation algorithm for inverse of convolution, Inverse-Flow, and Affine-StableSR. The second part applies generative models to agricultural seed quality assessment, unsupervised geological mapping, privacy-preserving anonymization of driving datasets, and art restoration.

PhD thesis, IIIT-Hyderabad