Lihe Ding

I am a final-year Ph.D. candidate from MMLab, The Chinese University of Hong Kong, and my advisor is Prof. Tianfan Xue.

I received both my Bachelor's and Master's degrees from Beijing Institute of Technology, supervised by Prof. Jianan Li. I interned at Kling AI, Kuaishou Technology, and Galbot. I also spent a wonderful time in Qcraft and SenseTime as a research intern, mentored by Boyin Zhang and Dr. Zhanpeng Huang, respectively.

CV  /  Google Scholar  /  Twitter  /  Github
Email: dean.dinglihe at outlook dot com

I have been working on world model, embodied AI, neural rendering, and 3D generation. My previous work also included 3D perception on Point Clouds. (*: Equal Contribution †: Corresponding Author)

We systematically evaluate GPT-6 Astra as an embodied policy across six domains, revealing a gap between useful task decisions and reliable physical control.

We introduce RoboSynChallenge, a competition that couples large-scale synthetic data generation with standardized real-world evaluation of generalizable manipulation policies.

We present a real-time closed-loop planning-control framework for reactive humanoid whole-body control, including zero-shot moving-target reaching on Unitree G1.

We present a novel 4D VAE that operates directly in native 4D space, that is dynamic colored voxel space, without 2D projection.

Fullpart generates each 3d part at full resolution. We also present PartVerse-XL, the largest human annotated 3d part dataset.

Copart generates 3d parts from contextual part latents and supports various application, such as articulation modeling.

LoRA-Edit enables controllable video editing through mask-aware LoRA fine-tuning.

we propose DynamicTree, the first framework that can generate long-term, interactive animation of 3D Gaussian Splatting trees.

We propose a hybrid pipeline for inserting 3D objects into videos, combining 3D Gaussian Splatting rendering for temporal consistency and a 2D diffusion-based enhancement for photorealistic lighting.

We present a general, training-free optimization method that achieves high-fidelity generation in diffusion models.

Interactive3D grants users precise control over the generative process through extensive 3D interaction capabilities.

We intergrate both 3D and 2D diffusion models with powerful priors into a unified framework with bidirectional guidance.

We propose Obj-NeRF, a novel comprehensive pipeline to extract object NeRFs from multi-view images, which can be used for NeRF editing tasks.

we propose an alternative to make sample-adaptive transformations based on the structure of the sample to cope with potential corruption via an auto-augmentation framework.

We establish new lidar-scanned scene flow datasets and propose a fast and hierarchical network for real-world scene flow estimation.

We propose a novel class-aware 3D proposal generation strategy and an efficient fully sparse convolutional 3D refinement module for vote-based Indoor 3D Detection.

We propose the first powerful 3D window-based transformer backbone on sparse 3D voxels leveraging mixed-scale information.

We propose a novel multi-modality two-stage approach to effectively and efficiently fuse point clouds and camera images in the Regions of Interest(RoI).

The dynamic graph constructing strategy can well capture the fine-grained geometry ignored by previous aggregation operations.

Template from JonBarron