Chunjiang Liu

Research Assistant
Carnegie Mellon University

Research Interests

Computer Vision Multi-Modal ML Generative Modeling 3D Vision

Contact

chunjiang.liu.richard@gmail.com

+1 (412) 616-4459

LinkedIn Profile

Google Scholar

Download CV

About

Chunjiang Liu is a Research Assistant at Carnegie Mellon University. He received his Master's degree in Electrical and Computer Engineering from CMU. His research focuses on advancing 3D Vision and Vision-Language Models. He is actively seeking PhD opportunities for Fall 2026.

Education

Master of Science in Electrical and Computer Engineering

Carnegie Mellon University, Pittsburgh, PA | Jan 2023 - Dec 2024

GPA: 3.83/4.00

Courses: Optimization, Advanced Computer Vision, Visual Learning & Recognition, Multi-modal ML, Deep Generative Modeling

Bachelor of Communication Engineering

University of Electronic Science and Technology of China (UESTC), Chengdu, China | Sep 2018 - Jun 2022

GPA: 3.84/4.00

Courses: Calculus, Linear Algebra, Probability and Statistics, Signal and System

Selected Publications

Total-Editing: Head Avatar with Editable Appearance, Motion, and Lighting

Yizhou Zhao, Chunjiang Liu, Haoyu Chen, Bhiksha Raj, Min Xu, Tadas Baltrusaitis, Mitch Rundle, HsiangTao Wu, Kamran Ghasedi

Accepted at ICCV 2025 Workshop E2E3D

ArXiv

Toward Material-Agnostic System Identification from Videos (MASIV)

Yizhou Zhao, Haoyu Chen, Chunjiang Liu, Zhenyang Li, Charles Herrmann, Junhwa Hur, Yinxiao Li, Ming-Hsuan Yang, Bhiksha Raj, Min Xu

Accepted at ICCV 2025 (Main Conference)

ArXiv Project Page

Work Experience

Innovation Center Intern

YAMAHA MOTOR SOLUTIONS CO., LTD. XIAMEN | Jul 2022 - Oct 2022

Built a dataset of industrial components with noisy backgrounds, enabling system robustness
Developed an unsupervised image segmentation algorithm based on mutual information maximization loss
Achieved strong visualization performance on the built dataset

Project Experience

Stage-Aware Vision-and-Language Navigation with Enhanced Auxiliary Loss

Course Project | Advisor: Daniel Fried | Fall 2024

Introduced a weighted angular distance loss to improve navigation decision granularity
Developed LLM-based instruction paraphrasing and instruction segmentation for data augmentation
Achieved a success rate (SR) of 53.2% on unseen environments on R2R dataset, outperforming strong baselines such as Speaker-Follower (2018), EnvDrop (2019) and ViL-CLIP (2021)

Improvement on Zero-Shot Text-to-Video Generation

Course Project | Advisor: Guila Fanti | Spring 2024

Investigated frame consistency problem in zero-shot video generation models
Proposed two enhancements: 1) LLM-based prompt elaboration, 2) Few-shot learning for frame consistency
Achieved improved CLIP scores (30.4 - ours vs 29.6 - baseline) on benchmark datasets

A Light-weighted Transformer for Next Word Prediction

Course Project | Advisor: Yuejie Chi | Spring 2024

Designed a lightweight transformer-based model for next-word prediction on edge devices like mobile phones
Applied knowledge distillation techniques, trained on Daily Dialog dataset (21,976 tokens)
Achieved a testing perplexity of 35 and demonstrated qualitative performance comparable to GPT-2