Visual Media Analysis and Generation Methods Based on Jittor

Lin Gao

Lingxiao Zhang

Jie Yang

Shuyu Chen

Fenglin Liu

Institute of Computing Technology, Chinese Academy of Sciences

Project Overview

About This Project

This project is based on the open-source machine learning framework Jittor, through which a comprehensive toolchain for visual media analysis and generation has been developed. A total of 20 algorithms have been open-sourced, including 5 integrated into the JGaussian. The toolchain primarily serves researchers and educators working on 3D model analysis and reconstruction, 3D editing and generation based on neural representations, and controllable digital human synthesis, as well as industry teams focused on digital humans and scene reconstruction by providing essential algorithmic tools.

The research work spans three major areas: geometric analysis and reconstruction, 3D editing and generation via neural representations, and controllable digital human synthesis.

This line of research has led to 6 publications in ACM TOG, 6 in IEEE TPAMI, 3 in IEEE TVCG, and 4 in top-tier conferences such as CVPR, NeurIPS, and ICCV. Several of these works have been featured on ACM TOG covers or highlighted in promotional trailers.

Geometric Analysis and Reconstruction

Neural Representation for 3D Editing and Generation

Controllable Digital Human Synthesis