Plenoxels and Voxel Radiance Fields
Plenoxels showed that you can reconstruct a 3D scene with NeRF-quality results without any neural network at all — just a grid of voxels storing color and density.
Overview
The result trains roughly 100x faster than the original NeRF while matching its visual quality.
Deep Dive
NeRF achieves photorealism but is slow because every sample requires a forward pass through a deep neural network, and training can take hours or days. Plenoxels (Sara Fridovich-Keil, Alex Yu et al., 2022) asked a provocative question: is the network even necessary? Their answer was no. They represent the scene as a sparse 3D voxel grid. Each occupied voxel stores a single opacity value plus spherical harmonic coefficients that encode view-dependent color. To render a pixel, the system trilinearly interpolates these values along the ray and composites them with standard volume rendering. Because there is no network, the whole thing is optimized directly with gradient descent on the voxel values, regularized for smoothness. The headline result: comparable quality to NeRF, trained in minutes on a single GPU.
Technical Insight
View-dependent color is the clever part. Instead of a network outputting RGB per viewing angle, each voxel stores a small set of spherical harmonic (SH) coefficients per color channel. Evaluating the SH basis in the ray's direction reconstructs how that point's color changes with viewpoint — capturing specular highlights and reflections. Opacity is direction-independent. Differentiable trilinear interpolation plus volume rendering makes every voxel value directly trainable, so optimization is a straightforward, network-free least-squares-style fit.
Strategic Impact
Speed and scale
Visual AI can automate inspection, detection, and tagging tasks at scale.
Build choices
Creative teams can prototype concepts faster with fewer manual revisions.
Team and workflow
Operations can use image and video signals that were previously hard to process.
The Future of Plenoxels and Voxel Radiance Fields
Plenoxels proved that the representation, not the neural network, drives NeRF's quality — a finding that reshaped the field. It directly inspired explicit and hybrid methods like Instant-NGP's hash grids and, ultimately, 3D Gaussian Splatting, which now dominates real-time radiance rendering. Expect continued movement toward explicit, GPU-friendly primitives that train in seconds and render in real time, with neural networks used selectively rather than as the core scene store.
Real-World Implementation
Quickly reconstructing a captured object into a 3D asset in minutes for e-commerce or museum digitization, instead of waiting hours.
Rapid prototyping of novel-view synthesis on a single consumer GPU for research and education.
Generating editable, explicit voxel scenes that artists can directly inspect and prune, unlike opaque network weights.
Serving as a teaching example that the scene representation, not deep learning, is what produces photorealistic results.
Risks & Guardrails
Image rights and consent can become legal risks if provenance is unclear.
Model performance can vary across lighting, demographics, and environments.
False positives may go unnoticed unless confidence thresholds are monitored.
Implementation Roadmap
Define acceptance criteria for precision, recall, and error costs.
Test with data that matches real production conditions.
Add human review for low-confidence or high-impact predictions.
Track model drift and revalidate after camera or dataset changes.
Keep Exploring
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Plenoxels and Voxel Radiance Fields quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Next guide
Neural Radiance Fields
Frequently asked questions
What is Plenoxels and Voxel Radiance Fields?
Plenoxels showed that you can reconstruct a 3D scene with NeRF-quality results without any neural network at all — just a grid of voxels storing color and density. The result trains roughly 100x faster than the original NeRF while matching its visual quality.
What is the most surprising design choice in Plenoxels compared to NeRF?
Plenoxels stores the scene directly in a voxel grid and optimizes those values, with no neural network in the pipeline.
What does each occupied voxel in Plenoxels store to capture view-dependent color?
Each voxel holds opacity plus spherical harmonic coefficients per color channel, which encode how color changes with viewing direction.
Roughly how much faster did Plenoxels train compared to the original NeRF?
By eliminating per-sample network passes, Plenoxels trained on the order of 100 times faster, finishing in minutes.
How are values fetched from the voxel grid along a ray?
Plenoxels uses differentiable trilinear interpolation to read color and opacity smoothly between voxel centers.
What broader conclusion did Plenoxels demonstrate to the field?
By matching NeRF quality with no network, Plenoxels showed the representation and optimization, not the MLP, were responsible for the results.