- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Multiresolution hash encodings such as Instant-NGP use a data-agnostic geometric resolution schedule and identical hash-table access at every spatial location, which the authors argue wastes capacity in low-information regions and induces collisions in information-dense ones. They recast hash-grid layout design as Retrievable Information Maximization (RIM), maximizing block–band empirical entropy discounted by a Poisson collision-survival factor, and propose RIM-Hash, which alternates an exact dynamic-programming update of per-level resolutions with a provably optimal update of block-wise soft gates that route between hash features and a trainable fallback feature. On Kodak, three gigapixel images, four WSIs, and seven Stanford SDFs, RIM-Hash reports a better parameter–fidelity trade-off than hash-grid and other INR baselines
Critical strengths
RIM-Hash shows large and consistent gains under a matched decoder, level count, and feature dimension across four signal types, including per-image wins on 22/24 Kodak images in PSNR.
The ablations separate the effect of each component. On Pluto, random or uniform gates with the same total gate mass fall 17 to 19 dB short of RIM gates, and the schedule and the gates help in complementary ways across budgets.
The learned gates behave as expected. They become denser as the hash table grows and sparser at fine levels, where they concentrate on edges and structures. This supports the claim that the surrogate captures meaningful signal structure
Critical weaknesses
The paper never reports the hash-grid configuration of the baselines or of the no-schedule ablation (minimum and maximum resolution, growth factor and per-level table size). This matters because, on Pluto, the ablation shows that the adaptive schedule drives most of the high-budget gain. RIM fixes the finest resolution to , so if the baselines’ finest resolution is below the image resolution, much of the headline gain may come from a capped baseline rather than from retrievable-information allocation.
The parameter–PSNR curve for Pluto does not match the reported numbers. In the curve, MetricGrids reaches about 67.4 dB at about 272M parameters, and this point is the basis of the “2.72× fewer parameters” claim. The numerical results instead report 62.49 dB at 271.98M parameters, and the qualitative comparison gives 63.51 dB at 238.8M. Since the main parameter-efficiency claims rest on these curves, it is unclear which values are correct
The paper shows that RI is a useful design heuristic, but not that it measures information that is actually retrievable in the trained model. Post-training collision rates and how close the gates are to binary are never measured.
The paper explains that its gains come from better allocation of fine-scale capacity. However, PSNR is largely driven by errors over wide smooth regions, and on SDFs the normal angular error, which reflects fine surface detail, improves only from 3.09 to 3.02. In addition, suppressed blocks use a trainable fallback vector shared within each level, and this vector can act like a learned intensity offset. Taken together, a substantial part of the PSNR gain may instead come from matching low-frequency tone and color more accurately.
Questions for authors
If possible, adding comparisons with ACORN, MINER and Compact NGP, which also allocate capacity adaptively or address hash collisions, would make the strengths of RIM-Hash stand out more clearly.
10 comments
Response to Reviewer nJWw
Post-training collision rates and how close the gates are to binary are never measured.
We thank the reviewer for this comment. We have added post-training collision rate and gate binarity analyses in Appendix R of the revised manuscript. Experiments on Pluto with show that the learned gates remain strongly concentrated near 0 and 1, indicating near-binary behavior. Moreover, the pre-training and post-training gate-weighted collision rates closely overlap across all 16 levels, with only minor deviations.
The paper explains that its gains come from better allocation of fine-scale capacity. However, PSNR is largely driven by errors over wide smooth regions,
PSNR is determined by the mean squared reconstruction error across all pixels and is not inherently dominated by wide smooth regions. As shown in the residual maps in Fig. 10 of revised paper, smooth regions exhibit substantially smaller reconstruction errors, whereas regions containing rich textures exhibit more residuals.
on SDFs the normal angular error, which reflects fine surface detail, improves only from 3.09 to 3.02.
We clarify that the primary objective of our original SDF experiments is to evaluate parameter efficiency rather than maximize reconstruction accuracy. Specifically, these experiments demonstrate that RIM-Hash achieves comparable geometric reconstruction fidelity with fewer parameters: 8.2% fewer than NeuRBF (786K vs. 856K) and 17.3% fewer than I-NGP and MetricGrids (786K vs. 950K).
To further assess reconstruction accuracy under a fair parameter-matched setting, we additionally scale RIM-Hash to 949K parameters, comparable to the 950K-parameter baselines. Under this setting, RIM-Hash reduces the average NAE from 3.09° to 2.98°, demonstrating that RIM-Hash is able to achieve better geometric reconstruction accuracy at a comparable parameter budget. These additional results have been included in the Table 16.
In addition, suppressed blocks use a trainable fallback vector shared within each level, and this vector can act like a learned intensity offset. Taken together, a substantial part of the PSNR gain may instead come from matching low-frequency tone and color more accurately.
To directly address this concern, we have added a frequency-resolved reconstruction error analysis on the Girl and Pluto images in Figs. 15 and 16 of Appendix O. Specifically, we decompose the reconstruction residuals of I-NGP and RIM-Hash into four spatial-frequency bands: 0–0.125, 0.125–0.25, 0.25–0.375, and 0.375–0.5 cycles per pixel. Panels (a)–(d) visualize the band-pass-filtered residuals, while panel (e) reports the corresponding MSE. The results show that RIM-Hash consistently achieves lower reconstruction errors than I-NGP across all four frequency bands. Importantly, the absolute MSE reductions in the intermediate-frequency bands (0.125–0.375 cycles per pixel) are substantially larger than those in the lowest-frequency band (0–0.125 cycles per pixel) on both images. This demonstrates that the reconstruction improvements do not mainly come from low-frequency tone and color.
In addition, the fallback-feature ablation in Table 12 shows that RIM-Hash maintains high reconstruction fidelity even when the trainable fallback vector is replaced with a fixed random feature, indicating that the improvements do not depend on learning the fallback representations. Taken together, the frequency-resolved analysis and fallback-feature ablation provide complementary evidence that the PSNR gains cannot be explained solely by learned fallback features or improved low-frequency intensity and color matching, supporting the effectiveness of RIM-Hash in allocating representational capacity for parameter-efficient signal fitting.
Response to Reviewer nJWw
If possible, adding comparisons with ACORN, MINER and Compact NGP, which also allocate capacity adaptively or address hash collisions, would make the strengths of RIM-Hash stand out more clearly.
We thank the reviewer for this constructive suggestion. In response, we have conducted additional experiments comparing RIM-Hash with ACORN and MINER on the gigapixel Pluto image, covering both low- and high-parameter regimes.
ACORN and MINER employ adaptive capacity allocation through architectures distinct from Instant-NGP-style multiresolution hash encoding. ACORN uses an adaptive hierarchical spatial decomposition to allocate resources according to local signal complexity, while MINER employs a Laplacian-pyramid-based multiscale representation with patch-wise MLPs. In contrast, RIM-Hash operates directly on multiresolution hash grids, explicitly accounting for collision-induced information loss to jointly optimize resolution scheduling and block-wise hash access.
As shown in the following table, RIM-Hash substantially outperforms both ACORN and MINER under both parameter budgets, demonstrating its superior parameter efficiency and reconstruction fidelity. These results further highlight the effectiveness of our retrievable-information-based capacity allocation strategy compared with alternative adaptive neural representations. We include these numerical results into the revised paper (Table 8).
| Method | Params (M) | PSNR (dB) | SSIM |
|---|---|---|---|
| ACORN | 8.015 | 35.90 | 0.9362 |
| ACORN | 103.73 | 36.34 | 0.9408 |
| MINER | 8.48 | 34.48 | 0.9294 |
| MINER | 126.30 | 40.51 | 0.9683 |
| RIM-Hash | 7.15 | 42.08 | 0.9722 |
| RIM-Hash | 99.83 | 67.42 | 0.9999 |
Regarding Compact NGP, the authors have not publicly released its implementation. We are actively contacting the authors via an anonymous email address to request access to the source code.
https://github.com/nv-tlabs/compact-ngp
We have also updated the Related Work section to include discussions of ACORN, MINER, and Compact-NGP, further clarifying their distinctions from our proposed method.
Response to Reviewer nJWw
We sincerely thank the reviewer for the detailed and technically insightful assessment of our work. We particularly appreciate the reviewer's acknowledgment of (i) the substantial improvements achieved under matched experimental settings, and (ii) the meaningful spatial patterns exhibited by the learned gates. Beyond these encouraging observations.
The reviewer's constructive comments on Instant-NGP baseline design, the post-training behavior of learned gates, information actually retrievable from trained models, and the consistency of our experimental results have helped us substantially strengthen the rigor of our work. In response, we have conducted additional experiments and strengthened the corresponding analyses, as detailed below.
We have clarified the hash-grid configurations of the baselines in Appendix G of the revised manuscript. Specifically, we follow the original Instant-NGP configuration, using 16 levels, 2 features per level, and a base resolution of 16. For image fitting, the finest resolution is set to half the maximum image dimension. The geometric growth factor is computed as b=(NminNmax)1/(L−1).
We have revised the Experimental Setup and Wall-Clock Cost section in Appendix G to provide a detailed description of the resolution scheduling configurations used for the baselines.
To directly address this concern, we conducted additional experiments on the Pluto image, setting the finest resolution of I-NGP to match that of RIM-Hash, i.e., max(h,w)=8000. Then, we evaluated four hash-table sizes (log2T∈18,20,22,24) and five geometric resolution growth factors (b∈1.2,1.4,1.6,1.8,2.0). We have included these additional results in Figure 17, Appendix P of the revised manuscript.
As the reviewer anticipated, increasing the finest resolution and tuning the geometric growth factor indeed improve I-NGP's reconstruction performance. However, even after these adjustments, RIM-Hash consistently achieves substantially higher PSNR across all evaluated parameter budgets.
Thus, the substantial performance gains of RIM-Hash cannot be explained simply by the baseline's capped finest resolution. Even when I-NGP is allowed to reach the same finest resolution and its geometric schedule is tuned, RIM-Hash maintains a significant advantage, demonstrating the effectiveness of our proposed adaptive resolution scheduling in improving hash-grid parameter efficiency.
We thank the reviewer for carefully examining our results and identifying this typo. The reported PSNR value for MetricGrids in Table 8 was incorrect and should be 67.49 dB. We have corrected this error in the revised manuscript.
We thank the reviewer for raising this important question. To directly address this concern, we conducted experiments on Pluto using fully trained RIM-Hash models with three hash-table sizes (log2T∈18,20,22). Specifically, we measure information that is actually retrievable in the trained model by computing the empirically mutual information between quantized frequency bands of the original and reconstructed images, aggregated across all 16 levels for each spatial block.
As shown in the Figure 21 in the Appendix S, Predicted RI strongly correlates with empirically mutual information, achieving Pearson (r=0.98899) and Spearman (ρ=0.99644) across different spatial blocks and hash-table sizes. The empirically decoded information is generally higher than predicted RI, which is consistent with the additional representational capacity of decoder parameters.
Thus, RI effectively predicts the amount of information recovered by trained models, beyond merely serving as a design heuristic. We have included the experimental results and analysis in Appendix R of the revised manuscript.