Efficient and Dynamic-Step Deep Image Generation Models
Project Description
This project will focus on designing a novel and efficient algorithm for deep image generation. State-of-the-art generative models, particularly diffusion models, can create stunningly realistic images but often require a large number of iterative steps, making them computationally expensive and slow. The goal of this project is to develop a framework that can generate high-quality images in significantly fewer steps. We will explore a dynamic step allocation mechanism, which allows the model to adaptively use fewer steps for simpler image generation tasks and more steps for complex ones based on the initial noise prior or text conditions. This approach aims to strike an optimal balance between generation quality and computational efficiency, covering both unconditional and text-to-image generation scenarios.
Supervisor
TAN, Ping
Quota
1
Course type
UROP1000
UROP1100
Applicant's Roles
The applicant is expected to first conduct a literature review on recent advancements in efficient generative models. The primary responsibility will be to design, implement, and experiment with the proposed dynamic-step generation algorithm. This includes setting up experiments, training the model on large-scale image datasets, and systematically evaluating the performance in terms of both image quality (e.g., using FID score) and inference speed. The applicant will work closely with senior researchers to analyze results and iterate on the model architecture and training strategy.
Applicant's Learning Objectives
The applicant will gain a deep understanding of the theory behind modern deep generative models, especially diffusion models. They will acquire hands-on experience in implementing and training large-scale neural networks using deep learning frameworks like PyTorch. The student will learn the complete research pipeline, from formulating a research problem and designing a novel algorithm to conducting rigorous experiments and analyzing results. They will also gain valuable expertise in model efficiency and optimization, which are critical skills for deploying deep learning models in real-world applications.
Complexity of the project
Moderate