GeezerVison

Jing Tan (谈婧)

I'm a third-year Ph.D. Candidate in MMLab at The Chinese University of Hong Kong, supervised by Prof. Dahua Lin. I work on scene-level visual creation and manipulation using structured representations and multimodal generative models.

Previously, I obtained the B.Sc degree from Nanjing University in 2020, and the Master degree from Nanjing University under the supervision of Prof. Limin Wang in 2023.

I was an applied scientist intern at AWS AI Lab during the summer of 2025, where I collaborated with Prof. Zhuowen Tu and Prof. Jiajun Wu.

Email  /  CV  /  Scholar  /  Twitter  /  Github

Research (Google Scholar)

In my research, I consistently leverage the inherent structure of visual data. In video understanding, I exploit the natural temporal and semantic structure of videos to enable efficient perception. In visual scene generation, I study how structured scene representations support reliable and controllable creation and manipulation of visual content.

We show, in the context of geometric object transformations in scene images, that RL is more suitable than SFT to solve visual editing problems with verifiable rewards, in terms of data efficiency and editing accuracy.

Imagine360 lifts standard perspective video into 360-degree video with rich and structured motion, unlocking dynamic scene experience from full 360 degrees.

LayerPano3D generates full-view, explorable panoramic 3D scene from a single text prompt.

We present Temporal Perceiver (TP), a general architecture based on Transformer decoders as a unified solution to detect arbitrary generic boundaries, including shot-level, event-level and scene-level temporal boundaries.

PointTAD effectively tackles multi-label TAD by introducing a set of learnable query points to represent the action keyframes.

The first transformer-based framework for temporal action proposal generation.

SS4D is a native 4D generative model that synthesizes dynamic 3D objects directly from monocular video.

GenDoP is an auto-regressive model that generate artistic and expressive camera trajectories from text prompts and geometric cues.

IDArb is a Diffusion-based intrinsic decomposition framework for an arbitrary number of image inputs under varying illuminations.

Camera-controllable human image animation dataset and framework.

A new Dual-level query-based TAD framework to precisely detect actions from both instance-level and boundary-level.

Experience

• Conducted research in collaboration with Prof. Zhuowen Tu and Prof. Jiajun Wu on reinforcement learning–guided spatial-aware image editing, with Yantao Shen as the internship manager.
• Leveraged RLVR to enable precise geometric object transformations in images following text instructions.

• Worked on multi-label temporal action detection via learnable query points.
• Developed a more general video action detection framework capable of localizing specific actions among multiple simultaneous actions.

Selected Honors & Awards

• CSIG Master’s Thesis Incentive Program Awardee, China Society of Image and Graphics, 2025
• Outstanding Master’s Thesis Award (3/226), Nanjing University, 2024
• National Scholarship (6/226), Ministry of Education of China, 2022
• 3rd Place, GEBD-LOVEU Challenge @ CVPR Workshop, 2022

Academic Services

• Regular Conference reviewer for CVPR, ICCV, ECCV, NeurIPS, ICML, ICLR; Reviewer for SIGGRAPH 2025.
• Journal reviewer for IJCV and TIP.
• Teaching Assistant for IERG4998 and IERG4999 at CUHK.

Thanks Jon Barron for sharing the source code of this website template.