The name Neural Radiance Fields makes the network sound inseparable from the idea. Plenoxels asks a useful question: how much of the result comes from representing and rendering a radiance field, and how much requires a neural network?
The December Plenoxels paper uses a sparse voxel grid with spherical harmonics. Its parameters are optimized from calibrated images with gradient-based methods and regularization. The representation contains no neural component. The authors report substantially faster optimization than their NeRF baseline on the benchmark tasks they study.
Keep the renderer, change the storage
The important continuity is the rendering problem. A camera ray passes through the scene, encounters density and accumulates color. The training images still provide the target. What changes is where the scene information lives and how the renderer retrieves it.
A network computes an answer from its weights and the queried coordinates. An explicit grid stores values at spatial locations. Interpolation can produce values between those locations. These choices have different costs. A grid offers direct access but can consume substantial memory; a network distributes information through a learned function but requires computation for its queries.
Sparsity matters because much of a scene is empty. Allocating equally detailed storage throughout a large box wastes resources. A sparse representation keeps information where it is useful, though the data structure and traversal still have to manage that organization.
View-dependent appearance also needs a representation. Spherical harmonics provide a compact way to describe smoothly varying directional signals. In this setting, they let color vary with viewing direction rather than forcing each location to have one fixed color. The number of coefficients limits how complicated that variation can become.
Separate the idea from its implementation
I find this comparison valuable because it prevents the implementation from becoming the explanation. A successful result can depend on camera calibration, a differentiable renderer, a useful scene parameterization and an optimization objective. Calling the whole thing "neural" can make those contributions harder to see separately.
There is a practical lesson for tool builders. If two representations produce useful views, the right choice may depend on memory, training time, rendering cost and ease of inspection. The most compact mathematical description is not automatically the easiest format to deploy in an application.
It is also important to keep the task in view. Optimizing one captured scene is different from training a general model that predicts many unseen scenes. Removing the network from the former does not make every use of learned models unnecessary. It reveals something about the structure of this particular problem.
I would like to see this kind of comparison whenever a new graphics technique gains a strong identity. Keep the objective and the evidence, then ask which components are essential. Sometimes a simpler representation makes the useful idea easier to understand. Plenoxels gives a concrete reason to think of radiance fields as a family of scene representations, with neural networks as one way to store them.