Hanzhang Tu

Ph.D. candidate at Tsinghua University specializing in 3D/4D vision and neural rendering. Builds end-to-end research systems spanning data synthesis, model training, GPU optimization, and real-time deployment. Seeking a Research or Applied Scientist role in world models and 3D/4D vision.

PhD in Control Science and Engineering

Aug 2022 – Jun 2027 · Beijing, China

Department of Automation, Tsinghua University

Expected completion: June 2027. Advisor: Prof. Yebin Liu.

BEng in Measurement and Control Technology and Instrumentation

Aug 2018 – Jul 2022 · Beijing, China

Department of Precision Instrument, Tsinghua University

Minor in Computer Science.

Tele360: Real-Time Feed-Forward Human Reconstruction from Sparse Unposed Cameras

TOG / SIGGRAPH Asia 2026

Dynamic human reconstruction, freeview synthesis, real-time streaming

• Built a feed-forward pipeline that jointly estimates camera poses and reconstructs dynamic 3D Gaussians from 4–6 sparse, unposed 2K RGB streams.
• Delivered input-to-rendering at over 25 FPS on one consumer GPU, with Gaussian maps streamed to remote devices through standard video codecs.
• Iterative free-view rendering on various portable devices.

GBC-Splat: Generalizable Gaussian-Based Clothed Human Digitalization under Sparse RGB Cameras

CVPR 2025

Clothed-human digitalization, 3D Gaussians, sparse-view reconstruction

• Reconstructed complete clothed-human geometry and Gaussian appearance directly from sparse RGB input without subject-specific optimization.
• Combined implicit occupancy and cross-view disparity estimation to support loose clothing and photorealistic 6-DoF novel-view synthesis without an SMPL template.

Tele-Aloha: A Low-budget and High-authenticity Telepresence System Using Sparse RGB Cameras

SIGGRAPH 2024

Real-time telepresence, free-view synthesis, human performance rendering

• Designed an end-to-end capture, reconstruction, streaming, and autostereoscopic display pipeline using four cameras and one consumer GPU.
• Achieved 2048 × 2048 rendering at 30 FPS with end-to-end latency below 150 ms for bidirectional high-authenticity remote communication.

Tencent

Jul 2024 – Aug 2025

Large-model inference systems

• Implemented KV cache CPU offloading for mixed long- and short-generation workloads, improving large-model inference throughput on production data by 10%–230%.
• Extended the offloading path to tensor and pipeline parallelism for multi-GPU inference, broadening compatibility across distributed deployment configurations.
• Deployed the feature into the production codebase after rigorous cross-hardware validation.

3D Vision

3D/4D Reconstruction, Neural Rendering (NeRF, 3DGS), Multi-view Geometry, Dynamic Human/Scene Modeling, Geometric Foundation Model & World Models

Machine Learning

PyTorch, Triton, Multi-GPU Distributed Training, Synthetic Data Generation (Blender, 3ds Max), Feed-forward Architectures

Systems Deployment

GPU Optimization (Custom CUDA Kernels), Real-time Rendering (OpenGL, Metal), Inference Acceleration (TensorRT, ONNX), Low-latency Streaming, Multi-sensor Synchronization & Calibration

Programming & Tools

Python, C++, CUDA, Rust, Docker, Git

Tele360: Real-Time Feed-Forward Human Reconstruction from Sparse Unposed Cameras

Hanzhang Tu, Z. Liao, W. Min, J. Zhang, Y. Liu.

ACM Transactions on Graphics (SIGGRAPH Asia 2026 Journal Track), accepted.

Jointly estimates camera poses and predicts dynamic 3D Gaussians from sparse unposed streams, enabling real-time 2K input-to-rendering on one consumer GPU.

HADES: Human Avatar with Dynamic Explicit Hair Strands

Z. Liao, Hanzhang Tu, et al.

IEEE/CVF International Conference on Computer Vision (ICCV), 2025.

Represents hair as strands bound to 3D Gaussians and models inertial, velocity-aware motion.

GBC-Splat: Generalizable Gaussian-Based Clothed Human Digitalization under Sparse RGB Cameras

Hanzhang Tu, Z. Liao, et al.

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.

Reconstructs complete human shapes and Gaussian splats from sparse-view RGB inputs in a feed-forward manner, including loose garments.

GPS-Gaussian+: Generalizable Pixel-wise 3D Gaussian Splatting for Real-Time Human-Scene Rendering from Sparse Views

B. Zhou, S. Zheng, Hanzhang Tu, et al.

IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.

Directly regresses source-view Gaussian parameter maps for instant sparse-view novel-view synthesis without per-subject fine-tuning or optimization.

Tele-Aloha: A Low-budget and High-authenticity Telepresence System Using Sparse RGB Cameras

Hanzhang Tu, R. Shao, et al.

ACM SIGGRAPH, 2024.

Combines four RGB inputs, cascaded disparity estimation, 3D Gaussian neural rasterization, and weighted blending for real-time 2K telepresence.

Tensor4D: Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering

R. Shao, Z. Zheng, Hanzhang Tu, et al.

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR Highlight), 2023.

Hierarchically decomposes 4D spatiotemporal tensors into time-aware volumes and compact feature planes for memory-efficient dynamic scene modeling.

Full publication list: Google Scholar.

Reviewer for CVPR, ICCV, ECCV, NeurIPS, IEEE Transactions on Instrumentation and Measurement (TIM), and IEEE Transactions on Circuits and Systems for Video Technology (TCSVT).

Mandarin Chinese (native); English (CET-4: 625, CET-6: 594); Japanese (JLPT N2).