zju3dv/pointsplat

PointSplat is the research code released alongside a paper accepted at ECCV 2026, a major computer vision conference. The paper is titled Compact Gaussian Splatting via Human-Centric Prediction, and the project comes from researchers at Zhejiang University, ByteDance, and CUHK Shenzhen. The README itself does not explain what Gaussian splatting is in plain terms, so readers unfamiliar with the field would need outside background to fully follow it, though the general idea from the title and code is building compact 3D representations of people from images. The repository provides the code needed to run the trained model rather than a full explanation of the method. Setup starts with creating a Python 3.10 environment through conda and installing the listed dependencies. A pretrained checkpoint file can then be downloaded from Hugging Face using a single command, or manually from the same hosted location. Two example workflows are documented for running inference, meaning using the already trained model to process new data rather than training it from scratch. One uses example data from a dataset called DNA Rendering, and the other uses example data from THuman2.0, a human body scanning dataset. Both workflows involve downloading a zipped example data package from Hugging Face, unzipping it into a specific folder structure, and then running a Python script with CUDA, which requires an NVIDIA GPU. The scripts reference configuration files by name and point to the downloaded checkpoint and dataset locations. Links are provided to related repositories for readers who want access to the full, non-example versions of either dataset. The README ends with a citation entry for the associated paper, intended for other researchers who want to reference this work. Overall this project is aimed at people already working in 3D computer vision or human reconstruction research, since it assumes familiarity with concepts like Gaussian splatting, GPU accelerated inference, and the specific benchmark datasets it uses. It is written in Python.