Splattr

Efficiently reconstructing and rendering high-quality 3D scenes from 2D images in real time, without the heavy computational cost and latency of traditional methods like Neural Radiance Fields.

More concretely: traditional NeRF-style approaches are slow to train and render; real-time applications (AR/VR, simulations, interactive media) require low latency; maintaining photorealistic fidelity while improving performance is non-trivial; and handling large-scale scenes without exploding memory usage is difficult.

The problem becomes: how can we represent a 3D scene in a way that is both fast to render and visually accurate, while being scalable and trainable efficiently?

A production web application that converts video footage of rooms into interactive 3D point cloud models and reconstructed meshes using LongSplat — NVIDIA's state-of-the-art unposed 3D Gaussian splatting.

Representing a scene as a set of anisotropic 3D Gaussians (instead of dense neural fields) allows for real-time rendering with high fidelity by leveraging rasterization-friendly primitives and GPU acceleration.

Key assumptions in that approach: a scene can be approximated as volumetric Gaussians with position, covariance (shape and orientation), color, and opacity; those Gaussians can be projected ('splatted') onto the screen efficiently; sorting and blending Gaussians in screen space is much faster than volumetric ray marching; and optimization can still converge to photorealistic results with gradient-based methods.