Factorized Multi-Resolution HashGrid for Efficient Neural Radiance Fields: Execution on Edge-Devices
Authors: Kim Jun-Seong, Mingyu Kim, GeonU Kim, Tae-Hyun Oh, Jin-Hwa Kim
Organizations: Dept. of E.E., POSTECH · Dept. of CS, The Univ. of British Columbia · Grad. School of AI, POSTECH · School of Computing Engineering, KAIST · NAVER AI Lab, South Korea · AI Institute of Seoul Nat’l Univ.
We introduce Fact-Hash, a novel parameter-encoding method for training on-device neural radiance fields. Neural Radiance Fields (NeRF) have proven pivotal in 3D representations, but their applications are limited due to large computational resources. On-device training can open large application fields, providing strength in communication limitations, privacy concerns, and fast adaptation to a frequently changing scene. However, challenges such as limited resources (GPU memory, storage, and power) impede their deployment. To handle this, we introduce Fact-Hash, a novel parameter-encoding merging Tensor Factorization and Hash-encoding techniques. This integration offers two benefits: the use of rich high-resolution features and the few-shot robustness. In Fact-Hash, we project 3D coordinates into multiple lower-dimensional forms (2D or 1D) before applying the hash function and then aggregate them into a single feature. Comparative evaluations against state-of-the-art methods demonstrate Fact-Hash's superior memory efficiency, preserving quality and rendering speed. Fact-Hash saves memory usage by over one-third while maintaining the PSNR values compared to previous encoding methods. The on-device experiment validates the superiority of Fact-Hash compared to alternative positional encoding methods in computational efficiency and energy consumption. These findings highlight Fact-Hash as a promising solution to improve feature grid representation, address memory constraints, and improve quality in various applications. Project page: https://facthash.github.io/
Figures & tables
Fig. 1: Comparison of Instant-ngp [ 16 ] , TensoRF [ 14 ] , K-planes [ 15 ] and Ours in terms of PSNR and inference time on the edge-device. Training is conducted on a standard GPU machine, whereas the inference is performed on the edge-device aligning with standard edge-device utilization practices. The area of each circle represents the model size.
Fig. 2: Conceptual illustration of parameter-encoding; iNGP [ 16 ] , TensoRF [ 14 ] , K-planes [ 15 ] and the proposed method.
Fig. 3: Schematic of the proposed method, Fact-Hash. For a given point P , we first project the point into the tri-plane and obtain surrounding multi-resolution 2D indices of each plane. For the multi-resolution 2D indices, we lookup the corresponding features from the assigned hash tables, and perform bi-linear interpolation to compose 3 multi-resolution feature vectors fxy,fyz, and fzx . The interpolated feature vectors are multiplied with Hadamard product and decoded to the density σ and the color c with the shallow MLP decoders. We predict the color of ray using standard volumetric rendering formula with the densities and the colors along the ray and minimize the reconstruction loss with the regularization.
Case
Model
Avg. PSNR ↑
Avg. SSIM ↑
Avg. LPIPS ↓
Name
#Params (M) ↓
Size ↓ (Mb)
Full views
iNGP
11.46
45.24
31.59
0.953
0.061
TensoRF
17.38
68.69
32.88
0.961
0.028
K-planes
33.04
393.08
32.36
0.962
0.049
TensoRF
6.51
25.73
31.69
0.951
0.064
K-planes
16.53
64.82
32.06
0.956
0.057
TABLE I: Experimental results on the NeRF synthetic datasets.
Fig. 4: Qualitative results of 8 views case on the NeRF synthetic dataset. Rendered images are results of chair , hotdog cases in the NeRF synthetic dataset by iNGP, TensoRF, K-planes, and ours.
Fig. 5: PSNR (a) and SSIM (b) values according to the number of uniformly sampled inputs. All metrics are average value of 7 NeRF synthetic data.
Fig. 6: Visualization of bitfield on the NeRF synthetic dataset. Each case is trained on 8-views on chair , drums and hotdog .
Model
10 Watt ↓
20 Watt Avg. ↓
chair
drums
ficus
hotdog
lego
materials
mic
ship
Avg.
iNGP
71.8
142.3
507.0
167.4
126.3
251.3
96.9
566.6
241.2
191.4
TensoRF
109.5
199.8
174.3
199.2
188.4
395.6
149.0
641.1
257.1
219.6
K-Planes
97.8
195.8
159.1
206.4
167.8
339.3
122.1
504.7
224.1
185.9
Ours
93.7
164.5
145.1
130.7
161.8
319.1
89.6
229.8
166.8
133.2
TABLE II: rendering time for 200 images on the NeRF synthetic datasets.
Case
Model
Avg. PSNR ↑
Avg. SSIM ↑
Avg. LPIPS ↓
Name
#Params (M) ↓
Size ↓ (Mb)
Full views
iNGP
11.46
44.18
28.05
0.918
0.128
TensoRF
4.89
18.92
27.93
0.912
0.139
K-planes
16.53
63.30
28.40
0.923
0.120
Ours
3.40
13.22
28.42
0.919
0.121
10% views
iNGP
11.46
44.18
19.73
0.869
0.186
TABLE III: Experimental results on the tank and temples datasets.
Fig. 7: Qualitative results on the Tank and Temples for 10% training inputs.
Fig. 8: Qualitative results on the San Francisco Mission Bay dataset. For 1/3 views, we conduct the experiment with the default size of baselines [ 16 , 13 ] and use 1/3 number of images from the dataset for training. For Full views, we constrain the model sizes to ≈ 50 Mb and train in the full-view setting.
Fig. 9: Comparison between iNGP and Fact-Hash in various collision rate. PSNR values are averaged along the scenes where iNGP success in training, excluding hotdog, lego, materials .
Neural Radiance Fields (NeRF) have revolutionized novel view synthesis, yet their classical implementations remain computationally intensive for high-fidelity rendering. QNeRF recently demonstrated the feasibility of training NeRF on gate-based quantum computers by combining amplitude embedding, parameterized quantum circuits (PQCs), parity-based measurements, and volumetric rendering. However, QNeRF relies on classical sinusoidal positional encoding for spatial coordinates, which scales poorly with scene complexity and resolution. In this work, we replace the sinusoidal positional encoding for spatial coordinates with the multiresolution hash encoding from Instant-NGP while keeping the view-direction encoding, amplitude MLP, quantum circuit, parity measurement, output scaling, and volumetric rendering pipeline unchanged. This hybrid design, Hash-QNeRF, retains the quantum radiance prediction step while benefiting from the fast convergence and memory efficiency of learnable hash grids. On a synthetic Blender scene, we achieve a final training loss of 0.003534, corresponding to approximately 24.5 dB PSNR on the fitted batch. Noise resilience experiments using Qiskit FakeKyiv and FakeTorino backends yield state fidelities of 0.93 to 0.98, indicating that hash encoding does not degrade the quantum circuit's noise tolerance.
Neural Radiance Fields (NeRF) achieve high-quality novel-view synthesis, but their long training times and reliance on dense input views limit accessibility. We present a comparative study of three accelerated NeRF variants - DS-NeRF, TensoRF, and HashNeRF and explore extensions targeted at the low-compute, low-data regime. First, we add a depth-supervision loss derived from COLMAP keypoints to TensoRF (TensoRF-DS) and evaluate it on the LLFF dataset under reduced view counts. Second, we ablate the feature-decoding MLP of TensoRF and study the effect of input downsampling on PSNR and runtime on the synthetic Lego scene. Third, we propose four architectural variants of the HashNeRF color and density networks, including residual and convolutional designs, and report PSNR/training-time tradeoffs under matched iteration budgets. Under iso-time evaluation, none of our extensions conclusively outperform the published baselines, but the experiments characterize which extensions transfer to constrained settings and surface design questions for future work.
Alice Huang, Prathamesh Sonawane, Yashdeep Thorat +1
We introduce a probabilistic splat-based radiance field framework that retains the fast rasterization and test-time efficiency of 3D Gaussian Splatting (3DGS) while replacing heuristic primitive manipulation with gradient-based optimization of a volumetric probability density. Rather than relocating, splitting, or culling Gaussians via hand-tuned densification (e.g., ADC), we treat primitive locations as samples drawn from a persistent, learnable density. We instantiate this density using a novel, memory-efficient multi-scale hierarchical grid that enables end-to-end gradient-based optimization. To stabilize the optimization, we derive an unbiased gradient estimator with control variates that markedly reduces variance. By allowing probability mass to flow to where the loss demands, our framework eliminates brittle priors and naturally explores the volume, achieving state-of-the-art reconstruction quality on mip-NeRF 360 while preserving 3DGS-level rendering speed.
Mia Gaia Polansky, George Kopanas, Stephan Garbin +2