Author ORCID Identifier

https://orcid.org/0009-0004-4176-6923

Date of Award

Spring 6-4-2025

Document Type

Thesis (Undergraduate)

Department

Computer Science

First Advisor

Professor Yu-Wing Tai

Abstract

Text-to-video diffusion models rely heavily on powerful temporal modeling to produce coherent and high-quality video sequences. However, current approaches overwhelmingly rely on transformer-based self-attention, whose quadratic complexity with respect to token length poses a major computation bottleneck for high-resolution and/or long-duration video generations. To overcome this, we propose an efficient and expressive alternative based on kernelized linear attention with ELU-based feature maps. Our approach reduces complexity to linear in sequence length while preserving the modeling capacity of traditional attention. In addition, we introduce a progressive training strategy with positional encoding interpolation, enabling stable training across increasing sequence lengths without sacrificing convergence speed or generation quality. Together, these techniques form a scalable framework that integrates seamlessly with existing diffusion models. Experiments on standard text-to-video benchmarks show that our method achieves competitive or superior visual fidelity and temporal coherence, with lower memory usage and inference time. Our findings demonstrate that linear attention, combined with progressive temporal training, offers a practical and scalable foundation for efficient video diffusion.

Available for download on Friday, June 04, 2027

Share

COinS