zju3dv/InfiniSplat

InfiniSplat v1 & v2: Feed-Forward 3D Gaussian Reconstruction

What it solves

InfiniSplat provides a feed-forward method for reconstructing 3D Gaussian scenes from images, enabling novel-view rendering. It addresses the challenge of creating 3D representations from limited visual data, supporting both single-image and multi-image inputs.

How it works

The project is split into two versions:

  • v1: Focuses on monocular view synthesis, reconstructing 3D scenes from a single RGB image (with optional LiDAR depth support).
  • v2: Reconstructs 3D Gaussian scenes from multiple overlapping images, with or without calibrated camera poses.

Who it’s for

Researchers and developers working in 3D reconstruction, computer vision, and novel-view synthesis.

Highlights

  • Feed-forward reconstruction: Rapidly generates 3D Gaussians from images without needing per-scene optimization.
  • Flexible input: Supports single RGB images (v1) or multiple overlapping images (v2).
  • Depth integration: v1 allows for depth-sensor-guided reconstruction using LiDAR data.
  • Multi-view support: v2 handles multiple views with or without camera calibration.

Related

  • Project
    zju3dv/InfiniDepth

    InfiniDepth is a CVPR‑2026 project that turns a single RGB image (or RGB + sparse depth) into arbitrary‑resolution depth maps and 3D Gaussian‑splatting models. It provides ready‑to‑run inference scripts, a Hugging Face Gradio demo, multi‑view video processing, and a full training pipeline for fine‑tuning on custom data.

  • Project
    WebODM/OpenSplat

    OpenSplat is a C++ implementation of 3‑D Gaussian splatting that converts COLMAP/OpenSfM/ODX/nerfstudio data into compact Gaussian‑based scene files. It supports CUDA, ROCm (HIP), Apple Metal, and CPU‑only builds, offers Docker images, and provides a simple CLI for training, resuming, masking, and exporting to .ply/.splat/.spz/.rad formats.

  • Project
    playcanvas/supersplat-viewer

    SuperSplat Viewer is a browser‑based viewer for high‑density Gaussian‑splat 3‑D scenes. It ships as a static web app (or NPM package) that uses PlayCanvas’ engine with WebGPU/WebGL, supports many URL‑based configuration options, and offers debugging, AR/VR, and embedding capabilities.

  • Project
    inspatio/querysplat

    QuerySplat is a 3D Gaussian Splatting prediction framework that decouples geometry and appearance representations to improve the quality of 3D scene reconstructions from images.

  • Project
    ooolabdev/ooosplat

    OOOSplat is a cross‑platform desktop app that converts 360° video or image sequences into 3‑D Gaussian‑splatting models (PLY) entirely on the user’s computer. It bundles FFmpeg, COLMAP and the Brush trainer, offers three quality presets, lets you preview/edit the point cloud, and export MP4 or offline HTML. All processing stays local, with optional anonymous telemetry and a local AI‑agent (MCP) for task control.