Gaussian Splatting Gains Traction

- The underlying "splatting" technique dates back to the early 1990s but was computationally impractical until recent advancements in GPU hardware. The 2023 research from Inria revitalized the method, achieving real-time rendering speeds. - Unlike Neural Radiance Fields (NeRFs), which use a neural network to represent a scene implicitly, Gaussian Splatting uses an explicit representation of millions of 3D Gaussians, which can be thought of as colorful, transparent blobs. This explicit representation is a key factor in its faster rendering performance. - Training times for Gaussian Splatting are significantly shorter than for NeRFs, often completing in minutes rather than hours, while achieving comparable or superior visual quality. - For rendering, the original implementation of 3D Gaussian Splatting requires a CUDA-ready GPU with compute capability 7.0+ and recommends 24GB of VRAM for training high-quality models. However, workable results can be achieved on consumer-grade cards like the NVIDIA RTX 3060 (12GB). - The technology is being applied in diverse fields beyond visual effects, including autonomous driving simulations, digital twins for urban planning, and creating virtual tours of real estate. - While NeRFs produce compact models (10-50MB), Gaussian Splatting models are significantly larger, often requiring 500MB to 1.5GB per scene. - The technique excels at representing complex, non-solid objects like smoke, fire, and water, as well as thin structures that are challenging for traditional mesh-based photogrammetry. - Major software platforms are beginning to integrate Gaussian Splatting, with plugins and support available for game engines like Unreal Engine and web-based viewers using frameworks like Three.js.