CV - ZHANG, Zhiheng (張志恒)

Educational Background

Shandong University

Bachelor of Engineering in Artificial Intelligence

September 2021 – July 2025 (Expected)

GPA (100/100): 91.6/100
GPA Ranking: 3/48
University Info: Project 985 & 211 designated university, National Key University

Language Proficiency

Japanese JLPT N2
English TOEFL iBT 83

Awards & Scholarships

Academic Scholarship (First Class)
Outstanding Student Scholarship (Second Class)
International Exchange Scholarship
ICPC (International Collegiate Programming Contest) Shandong Provincial Contest - Silver Medal

Research & Academia

Undergraduate Thesis: Study on Line Art Generation and Colorization Using Diffusion Models

Built a diffusion-based system for sketch cleanup, staged line art colorization, and background stylization.
Leveraged LoRA, ControlNet, and multimodal prompts for flexible and guided generation.

Project 1: Survey on Knowledge Editing in Large Language Models

Studied and compared mainstream knowledge editing techniques including parameter-invariant and variant approaches.

Project 2: Smoke and Flame Detection via Semantic Segmentation

Implemented DeepLabV3-based model with optimized loss for imbalanced data.

Research Interests

Animation Production Support Technologies

Sequential drawing generation models
Automatic coloring based on color specifications
In-between frame generation
3D-to-2D cel-shading for hand-drawn animation style

International Exchange Experience

Japanese Language & Culture Program at Hosei University, Japan (June 2023, 2 weeks)

Undergraduate Thesis Abstract

Sketch Cleanup and Staged Colorization System Based on Diffusion Models

Background

Traditional text-to-image diffusion models often fail to align with artists' intent and workflow, especially in animation-style creation. Interviews with student artists indicate strong demand for automation in three key areas: sketch cleanup, colorization based on color hints and background generation.

This thesis proposes a modular AI-assisted drawing system tailored for these needs.

Method

A multi-stage pipeline is developed using diffusion models, with each stage enhanced by tailored ControlNet and LoRA configurations. The system supports: sketch cleanup from rough drawings; staged colorization including interactive color hinting and automatic prompt generation; and background synthesis with style adaptation and seamless blending.

Results