A2-Edit: Precise Reference-Guided Image Editing of Arbitrary Objects and Ambiguous Masks
Reference-guided object editing across categories, using only a coarse mask.
European Conference on Computer Vision (ECCV) 2026
Video generation & visual editing.
I’m a Ph.D. student at Shanghai Jiao Tong University, advised by Prof. Xiaohong Liu, and jointly trained at Shanghai Innovation Institute.
My research focuses on AIGC, especially video generation, video editing, and multimodal generation.
News
10 updates Scroll
-
Submitted 4 papers to AAAI 2027. Wish us luck!
-
Accepted A2-Edit at ECCV 2026.
-
Accepted SceneVLP in TCSVT.
-
Submitted 2 papers to NeurIPS 2026. Wish us luck!
-
Accepted LayerT2V at ICML 2026.
-
Released A2-Edit for reference-guided image restoration and editing. Submitted to ECCV 2026.
-
Released LayerT2V for multi-layer video generation, with a large-scale layered video dataset. Submitted to ICML 2026.
-
Accepted FlowDirector at CVPR 2026.
-
Released FlowDirector for training- and inversion-free video editing. Submitted to CVPR 2026.
-
Submitted 1 paper to TCSVT.
Education
2026.09 — Present
Shanghai Jiao Tong University
Ph.D. · Computer Science and Technology
2026.09 — Present
Shanghai Innovation Institute
Jointly Trained Ph.D. Student
2022.09 — 2026.06
Central South University
B.Eng. · Software Engineering
Publications & preprints
LayerT2V: A Unified Multi-Layer Video Generation Framework
Generates full videos, foregrounds, backgrounds, and alpha mattes together.
International Conference on Machine Learning (ICML) 2026
FlowDirector: Training-Free Flow Steering for Precise Text-to-Video Editing
Edits videos through direct flow steering, without training or inversion.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026
SceneVLP: Augmenting Vision-Language Pre-Training Models for Video Action Recognition with Video Scene Graphs
Uses video scene graphs to enrich action recognition with structured context.
IEEE Transactions on Circuits and Systems for Video Technology (TCSVT)
InsertFuse: A Unified Framework for Multi-Category Reference-Guided Image Insertion
Distills category-specific insertion experts into one model with spatial guidance.
arXiv:2608.06490
STEP-OPD: Rethinking Output Targets and Internal Dynamics in On-Policy Distillation for Diffusion Models
Combines beyond-teacher output targets with alignment of internal representation changes.
arXiv:2608.04887
From Redundancy to Efficiency: Adaptive Video Tokenization via Appearance-Motion Decoupling and Dynamic Token Allocation
Compresses appearance and motion separately, allocating tokens according to video complexity.
Preprint coming soon
RewardFlow: Learning to Align Flow Models with Reward Distributions
Matches reward distributions to retain multiple preferred generation trajectories.
Preprint coming soon