Invertible Diffusion Model Structure for Style Transfer Tasks

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Invertible Diffusion Model Structure for Style Transfer Tasks"

by

WANG Hesong

Abstract:

Prompt-based style transfer with diffusion models has gained immense 
popularity on mobile applications. It is the standard approach to double-store 
the source image rather than recover it on mobile devices, but this becomes 
challenging with the explosion of digital content. Previous style transfer 
methods introduce irreversible information losses, which are compounded by 
compression artifacts during image dissemination. To overcome this, we 
introduce DuRIC-Net, a novel framework for reversible style transfer that 
enables high-fidelity restoration of the source image even from compressed 
outputs. Building on invertible neural networks (INNs), we propose a 
dual-branch architecture that serves as an efficient image encoder. To address 
JPEG compression more effectively, we incorporate a differentiable JPEG proxy 
that facilitates accurate gradient propagation during training. Furthermore, 
our proposed Residue Coupling Layers and Auxiliary Recovery Network (ARNet) 
minimize restoration artifacts. We extensively evaluate DuRIC-Net on the 
OmniEdit-Filtered-1.2M [24] dataset and social-media images from various 
sources, ablating various INN subnet architectures. Results demonstrate that 
our model achieves the state-of-the-art performance in both approximation and 
restoration quality, with broad applications in robust, editable visual 
content creation.

Date            : 29 April 2026 (Wednesday)

Time            : 15:40 - 16:20

Venue           : Room 2126A (near Lift 19), HKUST

Advisor         : Dr. CHEN Qifeng

2nd Reader      : Prof. SANDER Pedro