Visionary is an open, web-native platform designed for real-time rendering of diverse Gaussian Splatting variants and 3D meshes directly within a browser. It addresses the challenge of deploying and comparing advanced neural rendering techniques by providing a unified, efficient, and accessible environment. The platform targets researchers, engineers, and power users working with world models and 3D graphics, offering a lightweight, "click-to-run" experience that significantly lowers the barrier to reproduction and deployment.

How It Works

Visionary employs a hybrid rendering architecture leveraging WebGPU for high-performance parallel sorting and rendering of millions of Gaussian particles. Core to its design is per-frame ONNX inference, enabling dynamic neural processing and the integration of plug-and-play algorithms through a standardized Gaussian Generator contract. This approach supports not only standard 3DGS but also MLP-based 3DGS, 4DGS, neural avatars, and custom generative models. The platform automatically handles depth compositing between Gaussian point clouds and standard meshes, resolving occlusion issues for complex scene compositions.

Quick Start & Requirements

• Install: Clone the repository (git clone https://github.com/Visionary-Laboratory/visionary.git), navigate into the directory (cd visionary), and install dependencies (npm install).
• Run: Start the development server using npm run dev.
• Prerequisites: Node.js (v18+ recommended) and a WebGPU-enabled browser (Chrome recommended).
• Hardware: A discrete GPU (NVIDIA/AMD) on Windows 10/11 is strongly recommended for stable performance.
• Links:
  • Online Editor: https://visionary-laboratory.github.io/visionary/index_visionary.html
  • Project Page: https://visionary-laboratory.github.io/visionary/
  • Paper: https://arxiv.org/abs/2512.08478
  • Video: https://youtu.be/-K8EjMfk09c
  • Documentation: https://ai4sports.opengvlab.com/help/index.html
  • Demo: http://localhost:3000/demo/simple/index.html (after running npm run dev)

Highlighted Details

• Native WebGPU rendering for efficient processing of millions of Gaussian particles.
• Hybrid rendering architecture supporting seamless integration of Gaussian point clouds and standard meshes.
• Universal asset loader supporting multiple formats including Static Gaussians (PLY, SPLAT, etc.), standard meshes (GLB, GLTF, etc.), and ONNX for 4DGS/Avatars/custom algorithms.

Maintenance & Community

The project is a collaborative effort involving Shanghai AI Laboratory, Sichuan University, The University of Tokyo, Shanghai Jiao Tong University, and Northwestern Polytechnical University. Specific community channels like Discord or Slack are not detailed in the README.

Licensing & Compatibility

The project is licensed under the Apache-2.0 license. This license is permissive and generally compatible with commercial use and closed-source linking.

Limitations & Caveats

Ubuntu is currently not supported due to a WebGPU bug affecting fp16 ONNX pipeline compatibility. macOS users may experience limited GPU performance on non-high-end chips (e.g., M4 Max or better), potentially leading to slow rendering or stuttering.