Unifies video generation and 3D world modeling by decoding video latents into explicit 4D Gaussians, enabling real-time rendering across viewpoints and timesteps.

Generates native 4D Gaussian Splatting from image and text conditions, using temporally structured latents to separate geometry from appearance.

A unified framework providing large-scale datasets, benchmarks, and generative models for UDF-based 3D shape representation and generation.

A 4D Gaussian Splatting framework using a 4-camera rig for high-fidelity dynamic scene reconstruction and rendering.

We present Uni3D, a unified and scalable 3D pretraining framework for large-scale 3D representation learning, and explore its limits at the scale of one billion parameters.

Introducing a powerful 3D generative model that generates Gaussian primitives in arbitrary numbers by functionally disentangling Gaussian Splatting.

We propose to assemble diverse deep priors from large models for scene generation from single images in a zero shot manner.

We present a fast learning framework capable of inferring signed distance functions from noisy shapes within one minute through noise-to-noise mapping.

UDiFF is a 3D diffusion model for unsigned distance fields (UDFs) which is capable to generate textured 3D shapes with open surfaces from text conditions or unconditionally.

We present CAP-UDF to represent shapes and scenes with arbitrary architecture by learning a Consistency-Aware unsigned distance function Progressively.

We present 3D-OAE, a novel self-supervised point cloud representation learning framework which is highly efficient and can be further transferred to various downstream tasks.

We design a triplet network to learn VoxelPoint-to-Pixel matching via a differentiable probabilistic PnP solver.

We propose to guide the learning of zero level set in UDF using the rest non-zero level sets via a projection procedure.

We present CAP-UDF to represent shapes and scenes with arbitrary architecture by learning a Consistency-Aware unsigned distance function Progressively.

Unifies video generation and 3D world modeling by decoding video latents into explicit 4D Gaussians, enabling real-time rendering across viewpoints and timesteps.

Reconstructs dynamic 4D Gaussian scenes from monocular videos in seconds, jointly modeling geometry, appearance, and motion for real-time novel view synthesis.

Generates native 4D Gaussian Splatting from image and text conditions, using temporally structured latents to separate geometry from appearance.

Gaussian splatting reconstruction with video diffusion priors for high-quality 3D scene recovery from sparse views or even single view.

A unified framework providing large-scale datasets, benchmarks, and generative models for UDF-based 3D shape representation and generation.

A 4D Gaussian Splatting framework using a 4-camera rig for high-fidelity dynamic scene reconstruction and rendering.

A geometry-aware framework that grows 3D Gaussians from point clouds with text guidance for controllable 3D generation.

Unsupervised point cloud denoising via consistency-aware Noise2Noise matching, requiring no clean point cloud supervision.

GAP Gaussianizes arbitrary point clouds with text guidance, fitting high-quality 3D Gaussian representations from raw inputs.

We present NeRFPrior, which adopts a neural radiance field as a prior to learn signed distance fields using volume rendering for indoor scene surface reconstruction.

Proposing an effective approach for smooth surface reconstruction from sparse point clouds.

We present Uni3D, a unified and scalable 3D pretraining framework for large-scale 3D representation learning, and explore its limits at the scale of one billion parameters.

Introducing a powerful 3D generative model that generates Gaussian primitives in arbitrary numbers by functionally disentangling Gaussian Splatting.

We propose to assemble diverse deep priors from large models for scene generation from single images in a zero shot manner.

We propose a sparse view 3D Gaussian Splatting framework which explores the self supervisions inherent in the binocular stereo consistency.

We present a fast learning framework capable of inferring signed distance functions from noisy shapes within one minute through noise-to-noise mapping.

UDiFF is a 3D diffusion model for unsigned distance fields (UDFs) which is capable to generate textured 3D shapes with open surfaces from text conditions or unconditionally.

We present CAP-UDF to represent shapes and scenes with arbitrary architecture by learning a Consistency-Aware unsigned distance function Progressively.

We present 3D-OAE, a novel self-supervised point cloud representation learning framework which is highly efficient and can be further transferred to various downstream tasks.

A research project on 3D vision at Jilin University that I served as the advisor.

We present a 3D generation method that incorporates explicit generalized 3D priors with 2D diffusion priors to enhance the capability of generating unambiguous 3D consistent geometries.

We propose to learn a Local Distance Indicator as local priors to guide arbitrary-scale point cloud upsampling.

We design a sparse view reconstruction framework that leverages on-surface priors from explicit points to achieve highly faithful surface reconstruction.

We design a triplet network to learn VoxelPoint-to-Pixel matching via a differentiable probabilistic PnP solver.

We propose to guide the learning of zero level set in UDF using the rest non-zero level sets via a projection procedure.

We propose a level set alignment loss to evaluate the parallelism of level sets, which can be minimized to achieve better gradient consistency and eliminate uncertainty in the field.

A research project on 3D vision at Jilin University that I served as the advisor.

We present NeAF, a novel schema to learn implicit angle fields for point normal estimation. NeAF extends the success of implicit functions (e.g. NeRF and DeepSDF) to normal estimation.

We present CAP-UDF to represent shapes and scenes with arbitrary architecture by learning a Consistency-Aware unsigned distance function Progressively.

We present 3DAttriFlow to obtain high-quality 2D-to-3D reconstruction and point cloud completion results, while allowing controlled semantic attribute editing.