The Diffusion Problem Solver with Dual-Phased Diffusion: Toward Universal Anime Generation

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

Final Year Thesis Oral Defense

Title: "The Diffusion Problem Solver with Dual-Phased Diffusion: Toward 
Universal Anime Generation"

by

PENG Xuelei

Abstract:

Current image generation frameworks are largely task-specific and cannot 
function as general solvers across diverse visual problems. In this thesis, 
we introduce the Diffusion Problem Solver (DPS), a training-free framework 
that interprets diffusion denoising as a dual-phased process characterized by 
high error in early stages and low error in later stages. This perspective 
enables a phased strategy that first stabilizes global structure and then 
refines local details. To achieve this, we propose a Predictive Drift 
Controller (PDC) that corrects identity drift in the mid-noise regime by 
stabilizing spatial features through feedback control, followed by Local 
Harmonic Field Guidance (LHFG), which formulates late-stage refinement as a 
Dirichlet boundary value problem solved via Jacobi relaxation on latent 
features. Together these mechanisms transform diffusion into a controllable 
problem-solving framework rather than a fixed generative pipeline. We 
demonstrate that DPS requires no additional training and achieves strong 
results across multiple anime-style generation tasks, including manga 
synthesis, virtual try-on, and image editing, suggesting a unified and 
generalizable approach to controllable image generation.


Date            : 27 April 2026 (Monday)

Time            : 14:00 - 14:40

Venue           : Room 2132C (near Lift 22), HKUST

Advisor         : Prof. TANG Chi-Keung

2nd Reader      : Prof. SANDER Pedro