Jiajian Xie

AI4GC Lab

Jiajian Xie

3D GenerationStep DistillationTraining-free Acceleration

About Me

Hi! I am Jiajian Xie, a master’s student at AI4GC Lab, Zhejiang University, advised by Shengyu Zhang. My research focuses on 3D content generation and efficient inference for visual generative models.

Specifically, I work on Speech-driven Talking Face & Text/Image-to-3D Generation, as well as Step Distillation & Training-free Acceleration Methods for Diffusion and Visual Autoregressive (VAR) models.

My goal is to make high-fidelity visual generation—spanning 3D, images, and video—more practical and efficient for real-world and real-time applications.

Research Interests

  • 3D Generation — Achieve controllable 3D synthesis by aligning the distributions between input modalities (e.g., text, speech, image) and 3D representations.
  • Diffusion Step Distillation — Training a few-step model to match the denoising trajectories and distribution shifts​ of the original multi-step diffusion process.
  • Training-free Acceleration — Identifying computational redundancy and reducing inference cost through dynamic skipping and approximate computation.

Education

  • 2024.09 - 2027.06 · M.S. in Artificial Intelligence, Zhejiang University.
  • 2020.09 - 2024.06 · B.S. in Electronic Business and Computer Science (Dual Degree), South China University of Technology.

Selected Papers

ICLR2025

EcoFace: Audio-Visual Emotional Co-Disentanglement Speech-Driven 3D Talking Face Generation

PaperProject

arXiv2025

ETC: training-free diffusion models acceleration with Error-aware Trend Consistency

Paper

arXiv2025

EC-Diff: Fast and High-Quality Edge-Cloud Collaborative Inference for Diffusion Models

Paper

Internships

  • 2026.05 - present, Tencent, TEG, Hunyuan3D, ShenZhen.
  • 2025.06 - 2025.12, Tencent, WXG, WeChat Vision, ShenZhen.