Crystal Kwong

Publication Date

Spring 2025

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Chris Pollett

Second Advisor

Mark Stamp

Third Advisor

Kevin Smith

Keywords

GAN (Generative adversarial network), StyleGAN2-ADA, Unconditional to conditional transfer, Quantization, ONNX, Web deployment, Skybox

Abstract

Generative adversarial networks (GANs) are known for their ability to generate high quality
images mimicking real life or even particular art styles. Yet for all their capability, casually
training a GAN on an average machine can be infeasible as GANs require an enormous amount
of time and data to train. Even with a trained GAN, model inference demands heavy
computations, making GANs difficult to deploy on applications. To address these limitations,
techniques such as transfer learning and quantization have been leveraged to speed up training of
GANs and lighten computational cost of GAN inference. This project aims to use such
techniques to efficiently train and optimize a sky image GAN for deployment on a skybox
generator web page. We conducted experiments which demonstrated that transfer learning, in the
form of direct weight splicing from a pre-trained unconditional model to a conditional model,
accelerates the conditional model’s training. After 100 thousands of images (kimg) of training,
the model with transferred weights achieved an FID score of 29.01, outperforming the
conditional model trained from scratch which obtained an FID score of 134.51. As for
quantization, our results appear less impressive as the GAN model size of 120 megabytes is
shrunk only to 117 megabytes, and inference speed seems to remain unaffected. The final web
page deploys this model through ONNX Runtime Web and presents an interface allowing users
to generate skyboxes based on cloud types, fulfilling a use case of a choice-based AI skybox
generator.

Download

Available for download on Monday, May 25, 2026

Share

COinS