zju3dv/InfiniSplat

[SIGGRAPH Asia 2026 Journal Track] InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

What it solves

InfiniSplat addresses the challenge of large-baseline monocular view synthesis, specifically the ability to reconstruct a 3D Gaussian Splatting (3DGS) scene from a single RGB image or a single RGB-depth pair.

How it works

The project implements an implicit Gaussian decoding approach. It supports two primary modes of reconstruction:

  • Monocular Reconstruction: Generates 3D Gaussians from a single RGB image.
  • Depth-Sensor-Guided Reconstruction: Uses a combination of an RGB image and depth data (such as from a LiDAR sensor) to guide the 3D reconstruction process.

Who it’s for

This tool is designed for researchers and developers working in 3D computer vision, view synthesis, and 3D scene reconstruction from limited visual data.

Highlights

  • Dual-Mode Input: Supports both RGB-only and depth-guided reconstruction.
  • 3DGS Output: Produces 3D Gaussian Splatting representations for efficient rendering.
  • Interactive Visualization: Includes a Gradio-based web demo with an interactive Gaussian viewer and options to download PLY files and HTML exports.
  • Academic Foundation: Accepted to SIGGRAPH Asia 2026 (Journal Track).

Related

  • Project
  • Project
  • Project
  • Project
  • Dispatch