Bake It Till You Make It: Ultrafast Spatial Texture-Atlas Splatting

Recent 3D Gaussian Splatting (3DGS) extensions have improved visual fidelity by using hash-grid-based appearance modeling, but these methods suffer from high computational overhead during rendering. This paper addresses the challenge of achieving high-quality, high-frequency texture reconstruction while maintaining the real-time performance necessary for latency-sensitive applications like AR/VR.

The authors propose a decoupled radiance representation that separates scene appearance into two components: a per-primitive view-dependent color and a view-independent high-frequency texture residual. During training, the geometry is modeled using 2D Beta surfels, while the texture residual is learned via a multiresolution spatial hash grid and a compact MLP. A key innovation is the 'baking' procedure: after training, the hash grid output is evaluated across the surfels and stored in a compressed RGB texture atlas. This allows the renderer to replace expensive neural field queries with fast, hardware-accelerated 2D texture lookups. The authors also introduce a falloff-reduction regularizer to minimize fragment overdraw and improve rendering speed.

By offloading high-frequency details to a baked texture atlas, the method achieves up to a fivefold speedup over standard 3DGS and significantly outperforms previous hybrid methods in rendering speed. The approach maintains state-of-the-art visual fidelity while enabling real-time 4K rendering at 60 FPS on consumer hardware. Furthermore, the use of BC7 block compression for the texture atlas reduces memory requirements to 1 byte per texel without noticeable quality loss, making the representation highly efficient for deployment.

This work bridges the gap between high-fidelity neural radiance fields and the performance requirements of real-time graphics. By eliminating the need for per-fragment neural network evaluations, the method makes complex, photorealistic 3D scenes accessible on edge devices, providing a scalable solution for interactive 3D applications.