Abstract
We present Lagrangian Hashing, a representation for neural fields combining the characteristics of fast training NeRF methods that rely on Eulerian grids (i.e.~InstantNGP), with those that employ points equipped with features as a way to represent information (e.g. 3D Gaussian Splatting or PointNeRF). We achieve this by incorporating a point-based representation into the high-resolution layers of the hierarchical hash tables of an InstantNGP representation. As our points are equipped with a field of influence, our representation can be interpreted as a mixture of Gaussians stored within the hash table. We propose a loss that encourages the movement of our Gaussians towards regions that require more representation budget to be sufficiently well represented. Our main finding is that our representation allows the reconstruction of signals using a more compact representation without compromising quality.
Introduction
In the rapidly evolving world of 3D vision, two "schools of thought" have dominated how we represent digital scenes: Eulerian methods (like InstantNGP), which use fixed grids to store data, and Lagrangian methods (like 3D Gaussian Splatting), which use moving points. While grids are fast to train, they are memory-hungry; while points are spatially adaptive, they often require complex data structures to manage.
Lagrangian Hashing (LagHash) is a hybrid representation that marries the best of both worlds to create highly compressed, high-quality neural fields. By storing "mini-mixtures" of Gaussians within hash buckets and using a novel guidance loss to move them toward surfaces, the method achieves state-of-the-art reconstruction quality with significantly fewer parameters than standard methods.
The Problem: Memory vs. Quality
Neural fields (like NeRFs) face a fundamental "memory-quality" tradeoff. Most high-performance models use feature grids, dividing space into a grid and storing data at the corners. However, these grids are usually uniform, allocating the same amount of memory to empty air as they do to the intricate details of a complex object.
While sparsity-based methods skip empty cells, they don't solve the problem of non-homogeneous data. A flat wall needs very few parameters, while a person's face needs many. To be truly efficient, a model needs to be spatially adaptive—moving its "representation budget" to where the complexity actually is.
Lagrangian Hashing
The core innovation of LagHash is the Eulerian-Lagrangian hybrid. The model starts with the hierarchical hash table structure popularized by InstantNGP for extremely fast lookups. In the high-resolution layers of this hierarchy, instead of storing a single static feature in each hash bucket, LagHash stores a small set of $K$ Gaussians, each with its own learnable position and feature.
When the model queries a point in space, it calculates a weighted average of the Gaussians stored in the nearby hash buckets based on their distance to the query point. To ensure the points don't just sit randomly in space, a guidance loss inspired by the Expectation-Maximization (EM) algorithm acts like a magnet, pulling the Gaussians toward the surfaces of the object where they are most needed to represent high-frequency details. By reusing existing hash infrastructure, LagHash avoids the expensive "nearest neighbor" searches typically required by point-based methods.
Key Results and Insights
The authors evaluated LagHash across 2D image fitting (Gigapixel) and 3D NeRF datasets (Synthetic Blender and Tanks & Temples). Key findings include:
- Superior Compression: In 2D image fitting, LagHash achieved comparable quality to InstantNGP while being 2.6x more compact. In 3D tasks, it matched state-of-the-art quality with roughly 1.8x fewer parameters.
- Spatial Adaptivity: Low-resolution layers are best left as standard grids (Eulerian) to capture general shape, while high-resolution layers benefit most from the point-based (Lagrangian) approach for fine details.
- Robustness: When memory is severely limited, LagHash significantly outperforms 3D Gaussian Splatting (3DGS), which suffers a sharp decline in quality when point counts are lowered.
- Visual Fidelity: The method is particularly effective at resolving "collisions" in hash tables that typically cause artifacts on smooth surfaces in other grid-based methods.
Conclusion
Lagrangian Hashing demonstrates that combining grid-based hashing with point-based adaptivity offers a powerful path toward efficient neural representations. While currently optimized for object-centric scenes, future work aims to extend this hybrid approach to unbounded environments and explore fully learnable Gaussian scales.
Summary updated