InterNeRF: Scaling Radiance Fields via Parameter Interpolation

TL;DR

InterNeRF introduces a scalable, out-of-core NeRF architecture that scales model capacity by interpolating a partitioned set of parameters based on camera origin. By using bilinear interpolation across a grid of parameter sets and loading only the relevant subsets during training and rendering, it achieves improved centimeter-scale geometry and texture in large multi-room scenes while keeping memory usage constrained. The method demonstrates strong gains over Zip-NeRF on large indoor datasets and highlights the trade-offs between grid resolution, training time, and capacity. This work enables higher-capacity NeRFs for expansive environments and points to practical paths for model parallelism and alternative interpolation schemes to further improve scalability and efficiency.

Abstract

Neural Radiance Fields (NeRFs) have unmatched fidelity on large, real-world scenes. A common approach for scaling NeRFs is to partition the scene into regions, each of which is assigned its own parameters. When implemented naively, such an approach is limited by poor test-time scaling and inconsistent appearance and geometry. We instead propose InterNeRF, a novel architecture for rendering a target view using a subset of the model's parameters. Our approach enables out-of-core training and rendering, increasing total model capacity with only a modest increase to training time. We demonstrate significant improvements in multi-room scenes while remaining competitive on standard benchmarks.

Figures (4)

• Figure 1: Our framework. (1) We partition the scene into a parameter grid and assign training cameras to each cell based on its origin. (2) For a given (training or target) view, we obtain mixing weights as the bilinear interpolation coefficients based on camera origin. (3) Each query point along the ray is used to index into a multi-resolution set of grid features per parameter set, with either explicit assignment or a hash table. The mixing weights are applied here to yield a single set of features. (4) Each parameter set also has its own set of MLP weights, which are combined using the same mixing weights to form a new MLP.
• Figure 2: Images from Berlin and NYC in the Zip-NeRF dataset, rendered by Zip-NeRF and InterNeRF with a 5$\times$4 grid.
• Figure 3: Renders of the garden scene in the mip-NeRF 360 dataset using Zip-NeRF and InterNeRF with a 3$\times$3 parameter grid.
• Figure 4: Additional qualitative results on the Zip-NeRF dataset. Our method consistently reconstructs centimeter-level texture detail in large, multi-room scenes.