acceptodds
Under review as a conference paper at ICLR 2027

RIM-Hash: Retrievable Information Maximization Based Multiresolution Hash Encoding for Parameter-Efficient Signal Fitting

Abstract

Multiresolution hash encodings, popularized by Instant-NGP, provide fast and compact implicit neural representations. However, their layout is data-agnostic, level resolutions follow a fixed geometric schedule, and every spatial location accesses the hash tables identically. This uniform capacity allocation is mismatched to natural signals, whose information varies sharply across space and frequency. Consequently, it wastes capacity on low-information regions while inducing hash collisions in information-dense ones. To improve the efficiency of the multiresolution hash encoding, we recast its layout design as a Retrievable Information Maximization (RIM) problem. Specifically, for a given hash-table size, RIM selects the layout that maximizes the information that remains retrievable after hash collisions. The resulting encoder, RIM-Hash, routes low-information blocks to lightweight fallback features and concentrates hash capacity where information is both present and recoverable. Across benchmarks involving 2D images and 3D signed-distance fields, RIM-Hash consistently achieves a better parameter–fidelity trade-off than strong hash-grid baselines. On Tokyo, RIM-Hash achieves fidelity comparable to the prior state-of-the-art method using only 11.43% of its parameters, while on Girl it improves PSNR by 12.43 dB at a comparable parameter scale. Code is available in the https://anonymous.4open.science/r/Retrievable-Information-Maximization-FCCD.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
INR, Information allocation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nJWwOpenReview ↗
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 max⁡(w,h)\max(w,h), 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
AuthorsOpenReview ↗

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.

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).

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=(Nmax⁡Nmin⁡)1/(L−1).b=\left(\frac{N_{\max}}{N_{\min}}\right)^{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.

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 max(h,w)max(h,w), 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.

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\max(h,w)=8000. Then, we evaluated four hash-table sizes (log⁡2T∈18,20,22,24\log_2 T \in {18,20,22,24}) and five geometric resolution growth factors (b∈1.2,1.4,1.6,1.8,2.0b \in {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.

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

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.

The paper shows that RI is a useful design heuristic, but not that it measures information that is actually retrievable in the trained model.

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 (log⁡2T∈18,20,22\log_2 T \in {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=0.98899r=0.98899) and Spearman (ρ=0.99644\rho=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.

AuthorsOpenReview ↗

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 log⁡2T={18,20,22}\log_2 T = \{18, 20, 22\} 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.

AuthorsOpenReview ↗

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).

MethodParams (M)PSNR (dB)SSIM
ACORN8.01535.900.9362
ACORN103.7336.340.9408
MINER8.4834.480.9294
MINER126.3040.510.9683
RIM-Hash7.1542.080.9722
RIM-Hash99.8367.420.9999

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.

Reviewer_56KfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes a method to fix some common issues found in multi-resolution hash encodings for implicit neural networks, which was first used by Instant-NGP. Authors point out that capacity allocation in existing methods neglects the nature of data and every special location access the hash table similarly. Authors argue that this is sub optimal and wastes capacity on areas of the signal with low information density while inducing hash-collisions. As a solution, authors frame the layout design as a Retrievable Information Maximization(RIM) problem. Authors then proceed to test their method over signal representation tasks such as Kodak image dataset, giga-pixel images, WSIs and signed distance functions.

Critical strengths

Novel and promising idea with a detailed theoretical foundation: The paper recasts multiresolution hash-grid design as a collision-aware information-allocation problem which is novel as to my knowledge. Paper presents theoretical explanations to support their arguments.

Computationally cheap and some good results: RIM-Hash changes only the encoder layout and adds a few seconds of preprocessing, so it fits alongside decoder improvements. It shows consistent improvements in experiments over hash-grid baselines across Kodak, gigapixel images, WSIs and 3D SDFs.

Paper is well written and shows good practice: They have provided a comprehensive evaluation of the proposed method with detailed ablations, they have released the code, and the limitations are identified and acknowledged.

Critical weaknesses

I think the biggest weakness is that, this method can only operate on known signals such as images, WSIs and SDFs. This constrains its real world applications to signal fitting. Inability for direct deployment in NeRF style inverse tasks is weakness. However, authors identify this weakness and report it in Appendix section(M).

Questions for authors

As I understood, in the high-budget regime several models have more parameters than the signal has values. For example, Pluto (8000×8000) has about 192M values if RGB, but the largest RIM model uses 302M parameters. At that size, the model can memorize the image rather than represent it efficiently. Any comments on this?

8 comments
AuthorsOpenReview ↗

Response to Reviewer 56Kf

We are grateful for the reviewer's encouraging evaluation of our work. We appreciate the reviewer for raising the important concern regarding overparameterization, which has helped us strengthen the rigor of our experimental evaluation and better clarify the motivation behind our high-budget experiments. Our detailed responses are provided below.

I think the biggest weakness is that, this method can only operate on known signals such as images, WSIs and SDFs. This constrains its real world applications to signal fitting. Inability for direct deployment in NeRF style inverse tasks is weakness. However, authors identify this weakness and report it in Appendix section(M).

We agree that the current formulation is limited to signal-fitting tasks. As explicitly stated in our title, “RIM-Hash: Retrievable Information Maximization Based Multiresolution Hash Encoding for Parameter-Efficient Signal Fitting,” our work focuses on improving the parameter efficiency of signal fitting through information-guided hash-grid allocation.

Extending RIM-Hash to NeRF presents a distinct challenge, as the underlying 3D scene is unknown, making it difficult to estimate the information represented at different spatial resolutions and allocate hash-grid capacity accordingly. Addressing this challenge requires estimating the spatial–frequency information of an unknown 3D scene from indirect observations, which is beyond the scope of this work. We have revised the limitations section to discuss this challenge more explicitly and highlight the extension of RIM-Hash to NeRF-based inverse rendering as an interesting direction for future research.

As I understood, in the high-budget regime several models have more parameters than the signal has values. For example, Pluto (8000×8000) has about 192M values if RGB, but the largest RIM model uses 302M parameters. At that size, the model can memorize the image rather than represent it efficiently. Any comments on this?

We thank the reviewer for raising this important point. We agree that, in the high-budget regime, the number of model parameters can exceed the number of signal values. Following the experimental protocol established in Fig. 4 of the original Instant-NGP paper [1], which evaluates gigapixel image fitting across a wide range of hash-table sizes, we conduct comparisons across the full parameter-budget range. This allows us to systematically evaluate the reconstruction fidelity of different multiresolution hash encodings under varying model capacities.

Beyond facilitating the fair comparison, the high-budget also provides a useful setting for investigating whether different INR methods can achieve near-lossless reconstruction when sufficient model capacity is available.

For the Pluto image, RIM-Hash achieves a PSNR exceeding 104 dB after quantizing the reconstructed image to 8-bit precision, indicating near-lossless signal recovery. In contrast, MetricGrids achieves only 67.53 dB under the same evaluation protocol, despite its substantial parameter budget, which demonstrates that simply increasing the number of parameters does not guarantee near-lossless reconstruction, and the allocation of model capacity across spatial regions and spectral resolutions remains critical, even in the overparameterized regime. By adapting the hash representation to the spatial and spectral information distribution of the target signal, RIM-Hash utilizes its available capacity more effectively than conventional multiresolution hash encodings.

We have clarified the motivation and interpretation of these high-budget experiments in the revised manuscript (Remark 1, page 8) and uploaded the corresponding experimental logs to our repository for independent verification.

[1] Müller, Thomas, et al. "Instant neural graphics primitives with a multiresolution hash encoding." ACM transactions on graphics (TOG) 41.4 (2022): 1-15.

AuthorsOpenReview ↗

Response to Reviewer 56Kf

We sincerely thank the reviewer for the encouraging assessment of our work and for recognizing its novelty, theoretical foundation, and experimental contributions.

We are particularly grateful for your insightful observations regarding the applicability of RIM-Hash and the interpretation of our high-budget experiments. Your constructive suggestions have motivated us to conduct additional analyses, strengthen the experimental rigor of our work, and clarify the significance of our findings.

We hope that our detailed responses, additional experimental evidence, and manuscript revisions have satisfactorily addressed your concerns. We would be very grateful if you could kindly confirm whether your concerns have been resolved. Should any questions or concerns remain, please do not hesitate to let us know. We would be more than happy to provide further clarification or additional evidence.

Thank you again for your valuable feedback and thoughtful consideration of our work.

Dear Reviewer 56Kf,

We sincerely thank you for your encouraging feedback, positive rating, and constructive suggestions. We greatly appreciate the time and effort you devoted to evaluating our work. We would be grateful if you could take a moment to review our responses and clarifications, and let us know whether they have adequately addressed your concerns. Your feedback is highly valuable to us, and we sincerely appreciate your continued consideration of our work.

Reviewer_2tXROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RIM-Hash, a signal-adaptive multiresolution hash-grid encoder for parameter-efficient implicit neural representations. The main idea is to formulate hash-grid layout design as a retrievable-information maximization problem. The method estimates block-wise frequency-band information, models hash collisions through a Poisson-occupancy-inspired survival factor, jointly selects a non-geometric resolution schedule and initializes block-wise gates that route features either through a hash grid or a shared fallback feature. The paper provides a level-wise gate optimization result and a dynamic-programming procedure for resolution scheduling. Experiments on Kodak images, gigapixel images, whole-slide images, and SDF fitting report improved parameter–fidelity trade-offs over several INR/hash-grid baselines.

Critical strengths
  1. Recasting multiresolution hash-grid layout selection as an information-allocation problem is a compelling perspective. In particular, explicitly incorporating spatial heterogeneity and collision effects is well motivated.

  2. The paper provides a concrete objective, an analytic characterization of the level-wise gate update, and a dynamic-programming formulation for the schedule update. The optimization procedure is more principled than simply hand-designing a resolution schedule.

  3. The evaluation includes standard-resolution images, large images, pathology WSIs, and 3D SDFs. The paper also contains multiple ablations on collision modeling, gate selection, fallback features, block size, and the number of candidate bands.

Critical weaknesses
  1. The SDF experiment reports only modest parameter reduction relative to the strongest baselines (e.g., 786K vs. 856K/950K), and the table appears to aggregate seven objects without reporting per-object results or variation. This makes it difficult to assess whether the method consistently generalizes beyond 2D image fitting. More detailed per-shape results, multiple parameter budgets, and sensitivity to 3D block size would help.

  2. The approach relies on assigning contiguous frequency bands to levels according to a Nyquist-style interpretation. However, trained hash-grid levels with nonlinear decoding and multilevel concatenation do not necessarily correspond to disjoint frequency bands. Figure 1 is suggestive, but the isolation procedure and the claim of spectral specialization need more rigorous evidence. It would help to compare against non-contiguous band assignments or to quantify the actual frequency response of trained levels.

  3. The paper reports initialization time and optimization time, but the gate representation can be large for gigapixel images and WSIs. The number of gate logits is proportional to the number of blocks times the number of levels. Although these are included in the parameter count, the paper should report the fraction of parameters devoted to gates, memory footprint during training/inference, and whether the gating operation affects random-access inference latency.

Questions for authors
  1. Are the large gains on Girl and Pluto reproducible across random seeds? Please report mean and standard deviation for representative budgets.

  2. In the gate-selection ablation, are gate logits subsequently trainable for the random-gate and uniform-gate variants? If yes, why do they not converge toward similarly useful allocations? If no, this comparison conflates the effect of initialization with the effect of allowing adaptive learned routing.

  3. How are the reported parameter counts computed when a selected resolution is below the hash-table capacity? Are unused hash-table entries excluded? Is the same accounting protocol used for all baselines?

15 comments
AuthorsOpenReview ↗

Response to Reviewer 2tXR

We greatly appreciate the reviewer's thoughtful assessment and constructive feedback. We are particularly grateful for the reviewer's positive recognition of (i) the information-theoretic perspective on hash-grid capacity allocation, (ii) the principled optimization framework, and (iii) the extensive experimental validation across diverse INR tasks. The reviewer's insightful suggestions have helped us further strengthen the manuscript. We address each concern in detail below.

The SDF experiment reports only modest parameter reduction relative to the strongest baselines (e.g., 786K vs. 856K/950K),

We acknowledge that parameter reduction in the SDF experiments is more modest than in image fitting. Specifically, RIM-Hash uses 786K parameters, representing an 8.2% reduction compared with NeuRBF (856K) and a 17.3% reduction compared with I-NGP and MetricGrids (950K), while maintaining comparable aggregate geometric fidelity. Although these improvements are not as significant as the gain on 2D signals, they demonstrate that the proposed RIM-Hash method remains effective for 3D SDF fitting.

and the table appears to aggregate seven objects without reporting per-object results or variation. This makes it difficult to assess whether the method consistently generalizes beyond 2D image fitting. More detailed per-shape results, multiple parameter budgets, and sensitivity to 3D block size would help.

To expose performance differences across shapes, the revised manuscript reports per-object IoU and NAE for all seven Stanford models in Table 16, making the gain across individual geometries visible. We additionally include a parameter-matched comparison (949K vs. 950K parameters), where RIM-Hash reduces the average NAE from 3.09 (MetricGrids) to 2.98, an improvement of 0.12, comparable in magnitude to the 0.10 improvement from NeuRBF to MetricGrids. The average IoU also improves from 0.99991 to 0.99994. Furthermore, we evaluate multiple parameter budgets (786K and 949K) and investigate sensitivity to 3D block size (83,163, and 3238^3, 16^3, \text{ and } 32^3), thereby providing a more comprehensive assessment of reconstruction consistency and parameter efficiency across 3D geometries.

AuthorsOpenReview ↗

Response to Reviewer 2tXR

The approach relies on assigning contiguous frequency bands to levels according to a Nyquist-style interpretation. However, trained hash-grid levels with nonlinear decoding and multilevel concatenation do not necessarily correspond to disjoint frequency bands. Figure 1 is suggestive, but the isolation procedure and the claim of spectral specialization need more rigorous evidence. It would help to compare against non-contiguous band assignments or to quantify the actual frequency response of trained levels.

Thank you for this comment. If we understand your concern correctly, it primarily addresses the frequency interpretation of the multiresolution hash encoding introduced by Instant-NGP, which forms the foundation of our approach and whose representation efficiency RIM-Hash aims to improve. Below, we clarify the underlying mechanism, the physical interpretation of individual hash-grid levels, and the role of our frequency-band approximation. ​ Instant-NGP represents a signal using multiple hash-grid levels with progressively increasing spatial resolutions. At each level ℓ\ell, a grid of resolution NℓN_\ell stores trainable feature vectors in a corresponding hash table. Given a spatial coordinate, the features at the surrounding grid vertices are retrieved and interpolated. The interpolated features from all levels are then concatenated and passed through a shared MLP decoder to reconstruct the signal. Thus, the increasing grid resolutions naturally establish a hierarchy of spatial scales. Specifically, level ℓ\ell has a spatial sampling interval proportional to 1/Nℓ1/N_\ell, which determines its Nyquist frequency. Since the resolutions increase monotonically, their corresponding Nyquist frequencies also increase monotonically. Thus, the physical role of the ℓ\ell-th hash table is to encode and represent the fine-grained information corresponding to the frequency band associated with NℓN_\ell. In particular, finer levels provide the capacity to represent fine-grained spatial variations that coarser grids cannot efficiently resolve.

Figure 1 explicitly visualizes the reconstructed signal obtained from each hash-grid level. Specifically, for a trained six-level Instant-NGP model, we retain the learned feature vectors of level ℓ\ell while setting all feature vectors in the other hash tables to the mean value of that level. We then evaluate the original MLP decoder to obtain the level-isolated reconstruction. As shown, the decoded image from each level mainly contributes to a certain frequency band, which can be regarded as the ’isolated frequency response of the trained levels’. We revised the text in Section 4 to make the claim more rigorous, and we added a revised Figure 1 to Appendix T to better visualize the reconstructed image from each isolated level.

AuthorsOpenReview ↗

Response to Reviewer 2tXR

The paper reports initialization time and optimization time, but the gate representation can be large for gigapixel images and WSIs. The number of gate logits is proportional to the number of blocks times the number of levels. Although these are included in the parameter count, the paper should report the fraction of parameters devoted to gates, memory footprint during training/inference, and whether the gating operation affects random-access inference latency.

We thank the reviewer for highlighting the computational and memory overhead introduced by the gating mechanism. To address this concern, we conducted additional profiling experiments on the Pluto image using an NVIDIA Tesla V100-32GB GPU with FP32 precision. We evaluated four hash-table configurations log⁡2T∈{18,20,22,24}\log_2 T \in \{18,20,22,24\}, comparing RIM-Hash with I-NGP under identical hardware and query conditions. All measurements were collected after three warm-up runs, followed by ten repetitions.

The results are summarized below, where each entry reports RIM-Hash / I-NGP. Memory measurements represent peak allocated GPU memory.

MetricMethod$\log_2 T = 18 $$\log_2 T = 20 $log⁡2T=22\log_2 T = 22 \log_2 T = 24 \
Training memory (MB)I-NGP4948.775332.776868.7713012.77
RIM-Hash4949.095333.096869.0913012.37
Random-access latency (ms)I-NGP45.12047.61551.95454.833
RIM-Hash47.74551.26357.45560.356
Random-access memory (MB)I-NGP4280.214376.214760.216296.21
RIM-Hash4416.314512.314896.316432.31
Full-image decoding (s)I-NGP2.8012.8352.8722.889
RIM-Hash3.0183.0773.1873.274
Full-image memory (MB)I-NGP490.25586.25970.252506.25
RIM-Hash498.52594.52978.522514.52

Inference Latency. To evaluate random-access inference, we independently and uniformly sampled 1,048,576 random pixel coordinates across the entire image and processed them in a single batch. Both methods used identical random coordinates and query orders. The results show that RIM-Hash introduces only 5.82%–10.59% additional random-access inference latency compared with I-NGP. For full-image decoding, we the full Pluto image, observing a latency increase of 7.8%–13.3%. These results demonstrate that the additional gating operations introduce modest computational overhead while preserving efficient random-access inference. Memory Consumption. RIM-Hash exhibits nearly identical peak allocated GPU memory during training, with differences below 0.4 MB across all four configurations. During random-access inference, the additional peak memory is approximately 136 MB (2.16%–3.18%), while full-image decoding requires only 8.27 MB (0.33%–1.69%) additional memory. These results indicate that the gating mechanism introduces limited memory overhead relative to the overall model execution.

Gate Parameter Overhead. Gate parameters are included in the reported parameter counts. In the following table, we report their fractions of the total model parameters for Tokyo, Girl, Pluto and 4 WSIs for the representative configurations, respectively. We include the table in the revised manuscript (Appendix Q).

Log table size18202224
Log table size#gate (M)# total (M)rate (%)# gate (M)# total (M)rate (%)# gate (M)# total (M)rate (%)# gate (M)# total (M)rate (%)
Pluto0.00647.150.08520.006427.220.02350.006499.830.00640.0064302.570.0021
Girl0.04727.150.66010.047228.230.16720.0472111.580.04230.0472423.860.0111
Tokyo0.12267.021.74640.122628.200.43480.1226103.000.11900.1226405.600.0302
WSI 10.10598.171.29620.105932.510.32570.105994.640.11190.1059462.590.0229
WSI 20.11897.891.50700.118932.520.36560.118992.760.12820.1189453.610.0262
WSI 30.05758.100.71110.057532.220.17850.057592.060.06250.0575449.800.0128
WSI 40.13827.561.82800.138232.220.42890.138292.150.15000.1382449.880.0307

Overall, these experiments demonstrate that RIM-Hash maintains comparable training memory consumption and efficient random-access inference, with modest additional computational and memory costs. We have included these profiling results in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.