💻 computer science

Factorized Multi-Resolution HashGrid for Efficient Neural Radiance Fields: Execution on Edge-Devices

The paper introduces Fact-Hash, a novel parameter-encoding method that combines tensor factorization and hash-encoding to enable efficient, high-quality on-device training of Neural Radiance Fields by significantly reducing memory usage and energy consumption while maintaining rendering speed and image quality.

Original authors: Kim Jun-Seong, Mingyu Kim, GeonU Kim, Tae-Hyun Oh, Jin-Hwa Kim

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kim Jun-Seong, Mingyu Kim, GeonU Kim, Tae-Hyun Oh, Jin-Hwa Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to create a perfect, 3D hologram of a room so you can walk around it virtually. In the past, doing this required a supercomputer the size of a refrigerator (a powerful GPU) because the "recipe" for the 3D world was massive and complicated.

This paper introduces a new method called Fact-Hash, designed to let your smartphone or a small drone (an "edge device") create these 3D holograms on its own, without needing a supercomputer or a constant internet connection.

Here is the breakdown using simple analogies:

The Problem: The "Heavy Suit"

Traditional 3D modeling (NeRF) is like trying to carry a library of encyclopedias in your pocket to describe a single room. It's too heavy.

Memory: It takes up too much space on your phone.
Battery: It drains your battery instantly.
Privacy: Usually, you have to send your photos to a big server to process them, which raises privacy concerns.

The Solution: Fact-Hash

The authors created a new way to "compress" the 3D data. They combined two existing techniques into one super-efficient method. Think of it like packing a suitcase for a trip.

1. The Two Old Packing Methods

Before this paper, people tried two main ways to pack the 3D data:

Method A (Hash-encoding): Imagine a giant, chaotic filing cabinet where you throw everything in. It's fast to find things, but sometimes you accidentally throw two different items into the same drawer (a "collision"). If you have too many items, the drawers get messy, and you lose details.
Method B (Tensor Factorization): Imagine breaking a big 3D cube into flat sheets (like a stack of paper). It's very organized and saves space, but if the room is huge, you need so many sheets that it becomes heavy again.

2. The New Method: Fact-Hash (The "Smart Folder")

Fact-Hash is the best of both worlds. It uses a clever trick called "Map and Reduce."

The Map (Flattening the World): Instead of trying to store the whole 3D room at once, Fact-Hash projects the 3D room onto three flat walls (like looking at the room from the top, the side, and the front). It turns a complex 3D puzzle into three simple 2D puzzles.
The Hash (The Magic Filing): It then takes these flat 2D puzzles and puts them into the "filing cabinet" (the hash table).
The Reduce (Putting it Back Together): Finally, it grabs the info from the three flat walls and multiplies them together to reconstruct the 3D object.

Why is this better?

Fewer Collisions: Because the data is flattened into 2D before being filed, items are less likely to get mixed up in the same drawer. It's like organizing books by color on a shelf (2D) rather than throwing them all in a single bin (3D).
Tiny Size: This method creates a "recipe" that is 30% smaller than previous methods. It's like shrinking a 500-page book down to 300 pages without losing the story.
Few-Shot Robustness: This is a fancy way of saying it works even if you only have a few photos. If you try to learn a room from just 8 photos, old methods get confused and make the room look blurry or weird. Fact-Hash stays sharp and clear, like a detective who can solve a crime with very few clues.

Real-World Impact

The authors tested this on a NVIDIA Jetson Xavier NX, which is a small computer used in drones and robots.

Speed: It renders (draws) the 3D view 40% faster than the competition.
Battery: It uses less power, meaning your drone can fly longer.
Privacy: You can train the 3D model right on your device. Your photos never leave your phone, keeping your data private.

The Bottom Line

Fact-Hash is like a new, ultra-efficient compression algorithm for 3D worlds. It allows small devices to build high-quality, detailed 3D maps of the world around them instantly, using less memory, less battery, and fewer photos than ever before. This opens the door for better AR glasses, autonomous drones, and privacy-focused 3D mapping on your phone.

1. Problem Statement

Neural Radiance Fields (NeRFs) have revolutionized 3D scene representation but face significant barriers to deployment on edge devices (e.g., drones, autonomous vehicles, mobile phones). The primary challenges include:

Resource Constraints: Edge devices have limited GPU memory, storage capacity, and power budgets. Traditional NeRFs and even recent efficient variants often exceed these limits (e.g., requiring >10GB for large scenes).
Training vs. Inference: While distillation and post-processing (quantization/pruning) can reduce model size for inference, they often require re-training or create discrepancies between the training and user models. There is a lack of methods that support direct on-device training efficiently.
Trade-offs in Existing Methods:
- Hash-encoding (e.g., Instant-NGP): Offers high resolution but suffers from hash collisions in 3D space, leading to instability and poor performance in few-shot scenarios (limited training data).
- Tensor Factorization (e.g., TensoRF, K-planes): Compresses parameters effectively but lacks the scalability of high-resolution features and can struggle with expressiveness in complex scenes.

2. Methodology: Fact-Hash

The authors propose Fact-Hash, a novel parameter-encoding method that harmoniously integrates Hash-encoding and Tensor Factorization (specifically Tri-plane decomposition) within a Map-Reduce framework.

Core Architecture

The method operates in two distinct stages:

Map Process (Projection):
- Instead of projecting 3D coordinates directly into a 3D hash grid, Fact-Hash projects the 3D point $P(x, y, z)$ onto three 2D planes (Tri-planes: $xy$, $yz$, $zx$).
- This reduces the dimensionality of the coordinate space before hashing.
Hashing & Interpolation:
- For each 2D projection, the method queries separate multi-resolution hash tables.
- Features are retrieved via bilinear interpolation from the hash tables associated with each plane.
Reduce Process (Aggregation):
- The three feature vectors ( $f_{xy}, f_{yz}, f_{zx}$ ) obtained from the 2D planes are aggregated into a single feature vector.
- Fact-Hash utilizes the Hadamard product (element-wise multiplication) for this aggregation, which preserves compactness and interaction between axes.
- The final feature vector is fed into a shallow MLP to predict color and density.

Key Technical Innovations

Collision Mitigation: By restricting hash collisions to 2D planes rather than the full 3D space, Fact-Hash significantly reduces the probability of feature collisions, leading to more stable learning, especially in few-shot scenarios.
Multi-Resolution: It employs multi-resolution hash grids for each plane, allowing the model to capture both coarse and fine details without a massive increase in parameter count.
Efficient Rendering Pipeline: The method incorporates rejection sampling (using bitfields or proposal networks) and sampling termination to skip empty space, accelerating rendering on resource-constrained hardware.

3. Key Contributions

Novel Encoding Scheme: Introduction of Fact-Hash, which merges the expressiveness of multi-resolution hashing with the parameter efficiency of tensor factorization.
On-Device Optimization: A method specifically designed for direct on-device training and inference, addressing strict memory and power constraints without requiring complex two-stage distillation pipelines.
Few-Shot Robustness: The method demonstrates superior stability and performance in scenarios with limited training data (sparse inputs) compared to state-of-the-art (SOTA) baselines.
Comprehensive Evaluation: Extensive experiments on synthetic datasets, real-world outdoor scenes (Tank and Temples), and large-scale urban reconstructions (San Francisco Mission Bay) on actual edge hardware (NVIDIA Jetson Xavier NX).

4. Experimental Results

The paper validates Fact-Hash against SOTA methods like Instant-NGP (iNGP), TensoRF, and K-planes.

Parameter Efficiency:
- Fact-Hash reduces model size by over 30% (saving >1/3 of memory) compared to previous methods while maintaining or improving quality.
- In the NeRF Synthetic dataset, Fact-Hash used only 3.40M parameters (13.54 MB) compared to iNGP's 11.46M (45.24 MB) and K-planes' 33.04M (393.08 MB).
Performance Quality:
- Full-View: Achieves competitive PSNR (e.g., 32.00 on NeRF Synthetic) comparable to larger models.
- Few-Shot (Sparse Inputs): Outperforms all baselines. In 8-view training scenarios, Fact-Hash achieved a PSNR of 23.68, significantly higher than iNGP (18.38) and K-planes (23.58), demonstrating robustness against overfitting.
Rendering Speed on Edge Devices:
- Tested on an NVIDIA Jetson Xavier NX (10W and 20W power settings).
- Fact-Hash achieved an average 40% faster rendering speed compared to re-implemented baselines.
- It is particularly faster in complex scenes (e.g., "ficus", "ship") where 3D grid methods (iNGP) require excessive sampling to avoid artifacts.
Real-World Applicability:
- Successfully reconstructed large-scale urban scenes (San Francisco Mission Bay) with constrained model sizes (~50MB), outperforming iNGP and K-planes in both full-view and few-shot settings.

5. Significance

Fact-Hash represents a significant step forward in making NeRFs practical for real-world edge applications.

Privacy & Latency: By enabling on-device training, it eliminates the need to upload sensitive data to the cloud, addressing privacy concerns and reducing communication latency.
Adaptability: Its robustness in few-shot scenarios allows for rapid adaptation to dynamic environments (e.g., changing road conditions for autonomous driving) without retraining massive models.
Scalability: The method proves that high-fidelity 3D reconstruction is feasible on consumer-grade hardware with strict power and memory limits, bridging the gap between theoretical 3D representation and practical deployment.

In conclusion, Fact-Hash successfully balances memory efficiency, computational speed, and reconstruction quality, establishing a new benchmark for on-device Neural Radiance Fields.