Visionary-Laboratory/visionary

What it solves

Visionary provides a high-performance, web-native way to render complex 3D scenes using Gaussian Splatting and traditional meshes. It eliminates the need for heavy software installations by allowing real-time rendering and neural processing directly in the browser, lowering the barrier for deploying and comparing 3DGS-family methods.

How it works

Built on WebGPU and ONNX Runtime, the platform uses a hybrid rendering architecture that handles depth mixing between Gaussian point clouds and standard 3D meshes. It employs a standardized "Gaussian Generator" contract, allowing it to run per-frame ONNX inference to generate or update Gaussians dynamically. This enables support for a wide range of assets, from static 3DGS to dynamic 4DGS and neural avatars, as well as feed-forward generative post-processing.

Who it’s for

It is designed for researchers and developers working with neural rendering, world models, and 3D scene reconstruction who want to deploy their models to the web with a concise TypeScript API and three.js integration.

Highlights

• WebGPU Powered: High-performance parallel sorting and rendering of millions of Gaussian particles.
• Hybrid Rendering: Seamless depth compositing between Gaussian splats and standard 3D meshes to solve occlusion issues.
• Universal Loader: Supports a vast array of formats including PLY, SPLAT, GLB, GLTF, FBX, OBJ, and ONNX-exported models.
• Extensible Architecture: Allows users to export their own custom algorithms to ONNX for plug-and-play rendering.