What is clipgstream?

ClipGStream is a research project published at CVPR 2026 that addresses a specific problem in 3D video reconstruction: how to build a high-quality 3D model of a moving scene recorded from multiple camera angles simultaneously, when the recording is long and the motion is complex. The technique it builds on is called Gaussian Splatting, which represents a 3D scene as a large collection of small, overlapping blobs whose position, size, color, and opacity are optimized to match what the cameras saw. This approach has become popular in research because it produces good results and renders quickly. The challenge with dynamic scenes, where things are moving, is that naive approaches process one video frame at a time, which leads to flickering between frames and becomes impractical for long recordings. ClipGStream's contribution is to process video in clips rather than individual frames. It divides the entire recording into short segments, trains a foundational representation of the scene from the first clip (called the Reference Clip), and then trains each subsequent clip by inheriting the static parts of the scene from that reference while only learning what changed. This inheritance across clips is what the paper claims prevents flickering artifacts between segments and makes it practical to handle recordings of arbitrary length. The repository provides training scripts, a small example dataset with 20 frames for quick testing, and instructions for preparing custom multi-view video datasets. Training requires a Linux machine with an NVIDIA GPU. The setup involves creating a Conda environment from a provided file and installing one additional GPU library separately. Training runs in two stages: first the Reference Clip, then each Source Clip independently, which allows using multiple GPUs in parallel for longer sequences. The project is affiliated with Peking University and Pengcheng Lab. The full README is longer than what was shown.