acceptodds
Under review as a conference paper at ICLR 2027

High-Fidelity and Topology-Lossless Lattice Compression via Learned Context Models

Abstract

Lattice structures have been widely used in various applications of additive manufacturing due to its superior physical properties and material efficiency. However, a lattice structure with a huge number of struts would consume massive memory and storage space, which hinders its practical deployment in large-scale design and manufacturing scenarios. The storage challenge becomes even more pressing when moving from individual lattices to assembled lattice models. To solve this issue, we propose the Context-based Lattice Compression (CLC) framework, which leverages learned context models to provide a storage-friendly lattice representation. Specifically, we construct two sets of multi-resolution learnable hash embeddings that separately encode the lattice nodes and struts, and then excavate context dependencies in the embeddings to enable probability prediction for information entropy reduction. Additionally, we exploit a context model to accurately estimate the probability of geometric residuals, which are crucial for high-fidelity restoration, and we also use Delta Encoding to compress the correction edge set for lossless topology reconstruction. We also utilize feature pruning and hash fusion to further improve compactness and context quality across levels. This design allows the framework to jointly address storage reduction, geometry restoration, and topology preservation within a unified compression pipeline. Quantitative and qualitative evaluations demonstrate that CLC achieves up to 16.58 compression over vanilla lattice representation, achieving a 1.95 size reduction over the SOTA lattice compression approach while simultaneously improving fidelity and topological integrity.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Lattice Compression, Learned Context Models, High-Fidelity Reconstruction, Hash Embeddings, Additive Manufacturing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_djs4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Context-based Lattice Compression (CLC), a framework for compressing large-scale lattice structures while preserving high geometric fidelity and lossless graph connectivity. The method employs multi-resolution hash embeddings to represent node coordinates, radii, and edge connectivity. Learned context models are used to estimate the probability distributions of hash embeddings and geometric residuals for entropy coding. For topology reconstruction, the method predicts edge connectivity from candidate edges and explicitly encodes correction edges to recover the original topology.

The authors further introduce feature pruning and hash fusion to improve compression efficiency. Experiments on three procedurally generated lattice datasets demonstrate improved compression ratios compared with existing baselines while maintaining low geometric reconstruction errors and exact topology recovery.

Critical strengths

S1. Effective Compression with High Geometric Fidelity and Exact Topology Recovery.

The method demonstrates strong compression performance on the evaluated datasets. In particular, CLC reduces the Tetgen lattice from 738.19 MB to 44.53 MB while achieving zero topology reconstruction error. Compared with UniMate, it achieves a smaller compressed size with improved geometric fidelity and topology preservation. The combination of learned geometric representation and explicit topology correction provides a practical solution for compressing large lattice graphs.

S2. Comprehensive Compression Pipeline.

The paper integrates geometric representation, topology prediction, learned entropy coding, and residual correction into a unified framework. The inclusion of feature pruning and hash fusion further improves compression efficiency. The ablation studies provide evidence that these components contribute to the reported compression performance.

Critical weaknesses

W1. Limited Generalizability and Evaluation.

The experiments are limited to three procedurally generated lattice datasets with relatively strong spatial regularities. The paper lacks evaluations on real-world industrial lattices or complex structures derived from arbitrary 3D meshes. Since the method heavily relies on spatial locality, its compression performance on more irregular geometries and diverse topologies remains unclear. The current experiments are insufficient to demonstrate its general applicability.

W2. Questionable Practical Utility and Unjustified Complexity.

The practical necessity of the proposed framework is unclear. Since all evaluated lattices are procedurally generated, storing the source geometry and generation parameters may be more efficient than compressing the resulting graphs. Moreover, CLC requires approximately 77 minutes for encoding and 25 minutes for decoding a large lattice, raising concerns about its practical usability. The paper also lacks comparisons with simpler geometry-aware compression methods, making it difficult to justify the substantial computational overhead of the learned architecture.

Questions for authors

Q1. Practical Motivation and Application Scenarios

What are the concrete application scenarios that require storing these large lattice graphs instead of regenerating them from source geometry and procedural parameters? Given the substantial decoding time, how would CLC benefit practical CAD or additive manufacturing workflows?

Q2. Compression vs. Regeneration

Since all evaluated datasets are procedurally generated, could the authors compare CLC against storing the source geometry and generation parameters, considering both storage requirements and reconstruction time?

Q3. Computational Cost vs. Compression Benefits

Given that CLC requires approximately 25 minutes to decode the Tetgen lattice, what practical advantages justify this computational cost compared with conventional compression methods that decode in seconds?

5 comments

We sincerely thank the reviewer for the thoughtful feedback and insightful questions.


[W1] Limited Generalizability and Evaluation.

[A] We thank the reviewer for raising this concern. Although the lattices are procedurally generated, some models based on mechanical CAD components, such as Gear, Hinge, and Connector, undergo topology optimization for different loading conditions. Graphics-derived models, including Dragon, Kitten, and Bunny, use varied generation parameters and incorporate random perturbations to node positions, radii, and connectivity. The lattices also include targeted local refinements, including near boundaries or voids. We have added two tables and a figure in Appendix D.4: the first summarizes the lattice models, the second reports geometric variability and topological characteristics for nine instances, and the figure shows their node-degree distributions. These measurements document variation across the characterized Tetgen, Voronoi, and Hexahedron instances. For the reviewer's convenience, we reproduce the two tables below.

Table 1. Statistics of the generated lattice models.

Source ModelTypeBox size (mm)Num. nodesNum. strutsMin. node radius (µm)Max. node radius (µm)Min. strut len. (µm)Max. strut len. (µm)
GearTetgen50 × 50 × 1310,995,06674,765,9522.07129.098.44406.47
BootTetgen25 × 50 × 2512,637,07275,176,7791.34122.246.53391.98
DragonTetgen15 × 50 × 812,521,92874,635,3042.19131.1312.24504.14
MoldTetgen17 × 13 × 5011,025,96354,734,9402.35132.007.78459.53
ScrewTetgen18 × 16 × 5010,990,87475,239,2782.06147.669.42453.46
KittenVoronoi22 × 17 × 5033,095,45165,936,0080.7595.953.82290.27
ImpellerVoronoi50 × 50 × 1933,235,84058,989,9200.89104.354.33279.04
HingeVoronoi50 × 31 × 1933,239,23260,052,9280.6190.603.62354.11
TeddyVoronoi14 × 14 × 5033,192,03658,935,0120.78120.485.78290.55
BevelGearVoronoi20 × 20 × 5033,682,56362,528,3731.01116.833.53275.06
CrankHexahedron50 × 50 × 4511,413,99935,529,9001.7317.3110.23553.96
ConnectorHexahedron50 × 34 × 3111,475,27531,102,3252.0119.8817.00630.24
FlangeHexahedron50 × 50 × 1011,453,57230,528,8471.8517.5612.53581.10
BunnyHexahedron34 × 44 × 5011,486,29031,392,1351.2617.959.07608.39
BracketHexahedron50 × 50 × 2511,451,13629,367,5382.1719.6010.56563.67

Table 2. Geometric variability and topological characteristics of nine lattice models. CV is the coefficient of variation, degree entropy uses base-2 logarithms, and edge-orientation anisotropy is (3λₘₐₓ − 1)/2, where λₘₐₓ is the largest eigenvalue of the mean outer product of nonzero unit-edge directions.

MetricGearDragonMoldKittenImpellerHingeCrankConnectorBracket
Mean node degree13.599911.92079.92843.98463.54983.61346.22575.42085.1292
Degree entropy (bits)2.19063.54583.16780.11491.10871.07002.59131.64431.7881
Node-radius CV0.51560.71340.18020.57120.87790.75860.47240.16260.4727
Nearest-neighbor distance CV0.34010.55070.23150.47760.67810.61771.29090.35040.2821
Edge-length CV0.37150.64800.39360.44110.60570.54690.99230.36760.3338
Edge-orientation anisotropy0.00840.01150.01760.00030.02160.00910.02090.00920.0215

[W2] Questionable Practical Utility and Unjustified Complexity.

[A] We thank the reviewer for raising these practical concerns. Procedural generation helps us construct varied lattice instances for evaluation. CAD-based models such as Gear, Hinge, and Connector undergo topology optimization for different loading conditions. Graphics-derived models such as Dragon, Kitten, and Bunny use different generation parameters and include random perturbations to node positions, radii, and connectivity. These modifications produce final lattice graphs that cannot be specified by the source shape and basic generation parameters alone. Different loading conditions and mechanical applications may also require distinct optimized structures. Table 1 in the main paper includes comparisons with the geometry-aware compressors Draco and G-PCC. We recognize that the current CLC implementation incurs substantial runtime and GPU-memory costs, as discussed in Appendix F. Improving the computational and memory efficiency of CLC remains an important direction for future work.


[Q1] Practical Motivation and Application Scenarios.

[A] We thank the reviewer for this question. For lattices adapted to specific loading conditions through topology optimization, assembled from different unit-cell types, or modified by stochastic changes to geometry and connectivity, the final node and edge graph is not determined by the source geometry and basic procedural parameters alone. In additive manufacturing, storing and exchanging the finalized graph can support design archiving and transfer to downstream preparation. The storage burden grows when multiple lattices are assembled into a larger model, making compact representation of the instantiated graph especially relevant. CLC reduces this storage burden while reconstructing geometry with high fidelity and preserving connectivity losslessly. Its current decoding time, however, limits its suitability for interactive CAD operations or time-critical workflows. We therefore see the immediate application as offline storage and exchange of finalized designs, with faster decoding needed for broader integration into design and manufacturing workflows. We have added this discussion to the second paragraph of the Conclusion.


[Q2] Compression vs. Regeneration.

[A] We thank the reviewer for suggesting this comparison. We agree that a source STL and basic parameters for generating a regular lattice can require only a few MB. These inputs, however, do not reproduce the final lattice graphs in our dataset. For example, CAD-based models such as Gear, Hinge, and Connector undergo topology optimization to meet different loading conditions. Graphics-derived models such as Dragon, Kitten, and Bunny are generated with varied parameter combinations and receive random perturbations to node positions, radii, and connectivity. The resulting structures are determined by these instance-specific design choices and modifications, not by the source geometry and basic generation parameters alone.


[Q3] Computational Cost vs. Compression Benefits.

[A] We agree that decoding the large Tetgen lattice takes approximately 25 minutes, whereas conventional codecs decode it in seconds. The relevant use case is offline storage and exchange of finalized lattice graphs, as discussed in Q1, rather than repeated decoding during interactive CAD editing. On Tetgen, Table 1 in the main paper reports a combined geometry-and-topology size of 55.99 MB for CLC at the highrate setting, compared with 187.82 MB for xz and 265.20 MB for zstd. This is a 3.35× and 4.74× reduction in stored size, respectively, with high-fidelity geometry and lossless topology. The storage advantage is useful when archiving or transferring finalized designs, particularly as lattice assemblies grow, but the current decoding time remains a limitation for latency-sensitive workflows.

Thank you for the detailed rebuttal and additional experiments. I acknowledge the new ablation results and clarifications.

However, my main concerns remain:

  1. Limited generalizability. Despite the additional statistics, the evaluation remains restricted to three procedurally generated lattice families, without convincing evidence of broader applicability.
  2. Questionable practical utility. The authors clarify the motivation for offline storage, but the substantial computational overhead remains a major limitation. The paper still lacks a quantitative comparison with procedural regeneration or a convincing real-world application demonstrating the benefits of this trade-off.

Overall, I appreciate the additional experiments, but they do not sufficiently address my fundamental concerns about the practical significance and general applicability of the method.

I maintain my original rating of 2 (Weak Reject).

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

This paper proposes a context-based lattice compression framework, which designs two sets of multi-resolution learnable hash embeddings for per-instance geometry and topology coding, respectively. A coarse-to-fine prediction scheme is constructed for hash feature coding. Feature pruning and hash fusion further improve the compactness and context quality of the hash features. In geometry coding, regular grid points are utilized for lattice nodes organization and hash grid construction. The resulting hash features support both the prediction of coordinate displacement and radius, as well as their residuals. In topology coding, the network performs similar hash-feature-based edge occupancy estimation, while the correction edge set is compressed by delta coding. Experimental results demonstrate that the proposed method achieves up to 16.58× compression ratio over vanilla lattice representation and achieve a 1.95× size reduction over the SOTA lattice compression approach with better fidelity and topological integrity.

Critical strengths
  1. The proposed network architecture is carefully tailored to the characteristics of lattice data. For geometry coding, regular grid points are used to organize lattice nodes and construct hash grids in place of ground-truth node coordinates. The hash features not only support geometric attribute prediction but also guide the entropy coding of corresponding residuals. For edge coding, an adaptive KNN-based heuristic search is introduced to identify candidate edges, thereby simplifying the network's edge prediction task.

  2. The proposed method achieves SOTA performance on various lattice data. Compared with the previous SOTA method UniMate, the proposed method achieves consistent performance gains and better topological preservation.

Critical weaknesses
  1. Substantial computational overhead. Appendix E.3 reports relatively long encoding and decoding times across multiple datasets. Constructing 3D hash grids also incurs substantial GPU memory overhead. Although the optimizations described in Appendices A.3 and A.4 partially address these issues, they come at the cost of additional training iterations.

  2. Operations similar to coarse-to-fine hash features coding and hash fusion have previously been used in CNC [1] (Chen et al., CVPR 2024). Feature pruning is also shown in Instant-NGP [2]. The authors should clarify how these components have been adapted or improved specifically for lattice data.

[1] Yihang Chen, Qianyi Wu, Mehrtash Harandi, and Jianfei Cai. How far can we compress instantngp-based nerf? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20321–20330, 2024. doi: 10.1109/CVPR52733.2024.01921.

[2] ThomasMuller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1–102:15, 2022. doi: 10.1145/3528223.3530127.

Questions for authors
  1. The effect of PCA transformation in establishing the point-to-point mapping between regular grid points and lattice nodes requires further explanation and supporting references.

  2. When regular grid points are used to construct the hash grids for geometry coding, how does the method ensure that neighboring regular grid points correspond to spatially neighboring lattice nodes? And do the pruning indices for the hash grid features need to be transmitted to the decoder?

  3. The sampling strategy in Section A.4 requires more detailed description. In particular, across which parts of the network are unsampled hash feature embeddings kept frozen?

2 comments

Thank you for your insightful feedback and thought-provoking questions.


[W1] Computational and memory overhead.

[A] We acknowledge that the current CLC implementation has substantial encoding and decoding costs and requires considerable GPU memory for hash-grid training. The strategies in Appendices A.3 and A.4 mitigate some of these costs but do not eliminate them; feature sampling may also require more iterations to cover all embeddings. We have stated these runtime and memory limitations in Appendix F. Improving the computational and memory efficiency of CLC remains an important direction for future work.


[W2] Lattice-specific hash coding and feature pruning.

[A] We acknowledge that our level-wise context models and hash fusion build on prior NeRF compression work. BiRF [1] uses binary hash features, while context-based NeRF compression (CNC) [2] models binary features with a Bernoulli distribution. However, restricting all hash features to binary values would limit the precision available for high-fidelity lattice geometry. We therefore introduce adaptive quantization for fine-level hash features and for coordinate and radius residuals. We additionally use Gaussian probability models to assign probabilities to the quantized values and define feature- and residual-entropy losses for training. Unlike fixed binary coding, the quantization step size and its learned refinement allow us to adjust the rate–distortion trade-off.

Instant-NGP [3] maintains an occupancy grid during training to skip empty regions during ray marching, this grid does not itself prune hash features. In our method, the lattice input positions are fixed, so we identify the required hash-grid entries once before training through voxel-vertex activation and fine-to-coarse propagation, and encode only these active features. This requires no occupancy grid updates.

References

[1] Seungjoo Shin and Jaesik Park. Binary radiance fields. Advances in neural information processing systems, 36:55919–55931, 2023.

[2] Yihang Chen, Qianyi Wu, Mehrtash Harandi, and Jianfei Cai. How far can we compress instant-ngp-based nerf? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20321–20330, 2024.

[3] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1–102:15, 2022.

[Q1] PCA alignment for grid-to-node mapping.

[A] We sincerely thank the reviewer for this insightful question. PCA provides the lattice centroid, principal directions, and axis-wise standard deviations [4, 5]. We rotate and translate the regular grid to the lattice's principal frame and scale its axes by the corresponding standard deviations, with a factor of √3 to account for the spread of grid coordinates in [−1,1][-1,1]. This reduces orientation and scale mismatch, for example when an elongated lattice is placed obliquely, before joint normalization and Hilbert sorting of the grid points and lattice nodes [6]. Pairing the sorted points establishes the bijection, PCA improves its initialization. We clarify this role in Appendix B.


[Q2] Spatial correspondence and pruning indices.

[A] PCA aligns the regular grid with the lattice before both point sets are normalized together and sorted by their Hilbert values. Spatially close points often have nearby Hilbert values, and neighboring positions in the resulting one-dimensional order tend to represent nearby 3D points [6]. We pair grid points and lattice nodes by their respective ranks in these orders. This provides a locality-preserving initialization for geometry learning, not a guarantee that every pair of neighboring grid points maps to neighboring nodes.

The hash-grid feature-pruning indices do not need to be transmitted for either geometry or topology decoding. Both sides use the same regular grid points for geometry and the same geometry-decoded node coordinates for topology. From these inputs, they apply identical voxel-vertex activation, fine-to-coarse propagation, and hash indexing, so the decoder reconstructs the pruning mask for each network. We have added this clarification to Appendix A.5.


[Q3] Fine-level sampling and unsampled hash features.

[A] We thank the reviewer for this question. At each context-model refresh, we select a randomly positioned subset covering approximately 20% of the entries at each fine hash level, separately for the geometric and topological networks. We gather the coarser-level embeddings associated with the sampled entries' voxel vertices and feed them to CxC_x or CeC_e to predict the sampled entries' probability parameters and quantization refinements. Unsampled entries reuse cached context predictions; their hash embeddings are not frozen in either network, since all valid embeddings remain involved in geometric or topological prediction and can receive gradients from the task and feature-entropy losses. We have clarified this distinction in Appendix A.4.

References

[4] Ian T. Jolliffe and Jorge Cadima. Principal component analysis: a review and recent developments. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374(2065):20150202, 2016.

[5] Feiran Li, Kent Fujiwara, Fumio Okura, and Yasuyuki Matsushita. A closer look at rotation-invariant deep point cloud analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16218–16227, 2021.

[6] Bongki Moon, H. V. Jagadish, Christos Faloutsos, and Joel H. Saltz. Analysis of the clustering properties of the hilbert space-filling curve. IEEE Transactions on Knowledge and Data Engineering, 13(1):124–141, 2001.

Reviewer_8aYoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Context-based Lattice Compression (CLC) for strut-based lattice graphs containing node coordinates, radii, and connectivity. The method uses separate multi-resolution hash representations for geometry and topology, together with coarse-to-fine context models for entropy coding. Quantized geometric residuals improve reconstruction fidelity, while a correction edge set enables exact topology recovery. Feature pruning and hash fusion further reduce storage. Experiments on three procedurally generated lattice types report substantial compression over the original representation and improvements over the evaluated baselines.

Critical strengths

Relevant application and coherent formulation. Compressing large lattice graphs while retaining accurate geometry and exact connectivity is practically meaningful. Separating geometric reconstruction from topology correction provides a reasonable framework for addressing these requirements.

Sound prediction-and-correction design. Explicitly encoding the symmetric difference between predicted and original edge sets provides a straightforward mechanism for exact topology recovery, provided that predictions are reproduced consistently during decoding.

Useful cost reporting and initial experiments. The paper accounts for learned parameters, residuals, and correction edges in the compressed representation. It includes several baseline categories, component ablations, and a runtime breakdown that exposes the computational costs.

Critical weaknesses

The methodological contribution appears incremental. Multi-resolution hash representations, coarse-to-fine context modeling, and hash fusion closely follow established ideas in context-based NeRF compression, which the paper acknowledges. The main advance appears to be adapting these techniques to lattice geometry and connectivity, combined with residual coding.

The ablations provide limited support for some central components. In Table 2, replacing the geometric-residual context model with fixed 8-bit residual coding changes geometric size from 30.67 MB to 30.75 MB, approximately a 0.26% difference at similar geometric accuracy. Removing both context modules with the 8+8-bit configuration increases total size from 44.53 MB to 48.17 MB, corresponding to approximately 7.56% savings for the full model.

The baseline selection and adaptation need stronger justification. UniMate is presented as the principal state-of-the-art lattice compression reference, whereas its original objectives concern metamaterial generation, property prediction, and condition confirmation. Its adaptation into a large-scale bitstream codec is insufficiently documented.

Important task-matched controls are also missing: geometric neighborhood prediction followed by the same correction-edge coding, direct Hilbert-ordered delta coding of the original edges, and hybrid codecs combining error-controlled geometry compression with lossless graph coding. These comparisons would clarify the value of the learned topology model.

Evaluation diversity and practical trade-offs remain unclear. The paper evaluates three generated topology families without clearly reporting the number and variability of independent lattice instances.

Questions for authors

What is the principal lattice-specific methodological advance over directly adapting an existing context-based hash compressor?

What compression performance is obtained without the learned topology network, using a geometric neighborhood predictor and the same correction-edge coding?

How are UniMate and UniTruss adapted into practical codecs for the evaluated large lattices?

How are current-level statistics, pruning information, and optimized candidate-search parameters made available to the decoder?

How many independent lattices are evaluated within each topology family?

6 comments

We sincerely thank the reviewer for the constructive feedback, which has greatly helped us refine and improve our work.


[W1] The methodological novelty relative to context-based NeRF compression appears incremental.

[A] We acknowledge that our level-wise context models and hash fusion build on prior NeRF compression work. BiRF [1] uses binary hash features, while context-based NeRF compression (CNC) [2] models binary features with a Bernoulli distribution. However, restricting all hash features to binary values would limit the precision available for high-fidelity lattice geometry. We therefore introduce adaptive quantization for fine-level hash features and for coordinate and radius residuals. We additionally use Gaussian probability models to assign probabilities to the quantized values and define feature- and residual-entropy losses for training. Unlike fixed binary coding, the quantization step size and its learned refinement allow us to adjust the rate–distortion trade-off.


[W2] Context-model ablations.

[A] We thank the reviewer for pointing out that the gains from the context models appear modest at the operating points shown in Table 2(a) of the main paper. The previously reported 8-bit residual-coding result differed slightly in geometric accuracy from the full model. In separate reruns, training fluctuations brought its geometric accuracy closer to that of the full model, while its GSIZE rose to 31.38 MB. Relative to the full model's 30.67 MB, this corresponds to a 2.31% increase in geometric size without the geometric-residual context model. We agree that this is a modest gain. The updated context-model ablation results are shown below.

Table 1. Context-model ablation.

MethodDMAXERMAXEGSIZETSIZE
Ours11.356.6330.6713.86
w/o LW (full)11.366.6564.9158.35
w/o LW (16)11.366.5545.7430.28
w/o LW (8)11.406.6633.8914.28
w/o LW (4)11.346.6131.3717.74
w/o GR (8)11.356.6331.3814.26
w/o GR (4)195.12104.3511.77248.53
w/o LW&GR (8+8)11.346.7134.0414.13
w/o LW&GR (4+8)11.386.7531.1517.69

[W3] Baseline selection and adaptation.

[A] We thank the reviewer for highlighting the need to explain our adaptations of UniMate and UniTruss. We agree that their original objectives differ from large-scale bitstream compression. We have added a detailed account in Appendix A.6 of how both methods are adapted to encode large lattices in patches, including their geometry and topology bitstreams and the treatment of cross-patch and correction edges.


[W4] Task-matched topology controls.

[A] We thank the reviewer for this suggestion. To assess the value of the learned topology network, we added an ablation in Appendix E.3 comparing our method with adaptive KNN followed by direct correction-edge coding and with direct coding of the original topology. As shown below, our topological size is smaller than both controls on all three datasets.

Table 2. Topological sizes in the network ablation.

DatasetOur TSIZEKNN TSIZEDirect TSIZEOriginal TSIZE
Tetgen12.59281.43121.27570.42
Voronoi56.59348.54102.33502.96
Hexahedron18.18265.1052.09271.07

References

[1] Seungjoo Shin and Jaesik Park. Binary radiance fields. Advances in neural information processing systems, 36:55919–55931, 2023.

[2] Yihang Chen, Qianyi Wu, Mehrtash Harandi, and Jianfei Cai. How far can we compress instant-ngp-based nerf? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20321–20330, 2024.

[W5] Dataset diversity and evaluation scope.

[A] We thank the reviewer for pointing out that the variability of the lattice datasets was not sufficiently documented. We have added two tables and a figure in Appendix D.4: the first summarizes the lattice models, the second reports geometric variability and topological characteristics for nine lattice instances, and the figure shows their node-degree distributions. These measurements document variation in geometry and topology across the characterized Tetgen, Voronoi, and Hexahedron instances. For the reviewer's convenience, we reproduce the two tables below.

Table 3. Statistics of the generated lattice models.

Source ModelTypeBox size (mm)Num. nodesNum. strutsMin. node radius (µm)Max. node radius (µm)Min. strut len. (µm)Max. strut len. (µm)
GearTetgen50 × 50 × 1310,995,06674,765,9522.07129.098.44406.47
BootTetgen25 × 50 × 2512,637,07275,176,7791.34122.246.53391.98
DragonTetgen15 × 50 × 812,521,92874,635,3042.19131.1312.24504.14
MoldTetgen17 × 13 × 5011,025,96354,734,9402.35132.007.78459.53
ScrewTetgen18 × 16 × 5010,990,87475,239,2782.06147.669.42453.46
KittenVoronoi22 × 17 × 5033,095,45165,936,0080.7595.953.82290.27
ImpellerVoronoi50 × 50 × 1933,235,84058,989,9200.89104.354.33279.04
HingeVoronoi50 × 31 × 1933,239,23260,052,9280.6190.603.62354.11
TeddyVoronoi14 × 14 × 5033,192,03658,935,0120.78120.485.78290.55
BevelGearVoronoi20 × 20 × 5033,682,56362,528,3731.01116.833.53275.06
CrankHexahedron50 × 50 × 4511,413,99935,529,9001.7317.3110.23553.96
ConnectorHexahedron50 × 34 × 3111,475,27531,102,3252.0119.8817.00630.24
FlangeHexahedron50 × 50 × 1011,453,57230,528,8471.8517.5612.53581.10
BunnyHexahedron34 × 44 × 5011,486,29031,392,1351.2617.959.07608.39
BracketHexahedron50 × 50 × 2511,451,13629,367,5382.1719.6010.56563.67

Table 4. Geometric variability and topological characteristics of nine lattice models. CV is the coefficient of variation, degree entropy uses base-2 logarithms, and edge-orientation anisotropy is (3λₘₐₓ − 1)/2, where λₘₐₓ is the largest eigenvalue of the mean outer product of nonzero unit-edge directions.

MetricGearDragonMoldKittenImpellerHingeCrankConnectorBracket
Mean node degree13.599911.92079.92843.98463.54983.61346.22575.42085.1292
Degree entropy (bits)2.19063.54583.16780.11491.10871.07002.59131.64431.7881
Node-radius CV0.51560.71340.18020.57120.87790.75860.47240.16260.4727
Nearest-neighbor distance CV0.34010.55070.23150.47760.67810.61771.29090.35040.2821
Edge-length CV0.37150.64800.39360.44110.60570.54690.99230.36760.3338
Edge-orientation anisotropy0.00840.01150.01760.00030.02160.00910.02090.00920.0215

[Q1] Lattice-specific advance over hash-based NeRF compression.

[A] As noted in our response to W1, we build on context-based hash compression rather than proposing a new hash-grid architecture. The lattice-specific advance is to encode high-fidelity node geometry together with a losslessly recoverable edge set. We adaptively quantize fine-level features and coordinate and radius residuals. Gaussian probability models yield trainable entropy losses for the quantized values, while adaptive step sizes control the rate–distortion trade-off. This differs from the binary features of BiRF [1] and the Bernoulli modeling of CNC [2]. For topology, we predict edges from restored coordinates and encode the correction edge set to recover the original connectivity exactly.


[Q2] Topology compression without the learned topology network.

[A] Please see our response to W4.


[Q3] UniMate and UniTruss codec adaptations.

[A] We describe the codec adaptations of UniMate and UniTruss in Appendix A.6. Both methods encode Hilbert-ordered 27-node patches and store cross-patch and correction edges. UniTruss codes quantized geometry, topology, and shared VAE latents together with geometric residuals. UniMate codes vector-quantized token indices and a codebook to reconstruct patch geometry, with geometric residuals and an adapted edge predictor for topology.


[Q4] Decoder-side statistics and search parameters.

[A] The current-level mean and standard deviation used for context coding are stored with each compressed network, so the decoder uses the same auxiliary statistics during entropy decoding. Feature-pruning masks, by contrast, are reconstructed rather than transmitted: both sides activate hash-grid vertices from the same regular grid points for geometry or reconstructed node coordinates for topology, including fine-to-coarse propagation. Candidate search uses eight optimized scalar parameters and a precomputed-neighbor limit, which are stored in the metadata. These nine search values have a small, fixed storage cost. We clarify these distinctions in Appendix A.5.


[Q5] Lattice instances per topology family.

[A] The primary compression comparison evaluates five lattice models from each of the Tetgen, Voronoi, and Hexahedron families and reports the mean of their results. As noted in our response to W5, Appendix D.4 summarizes the lattice models and characterizes lattice instances through geometric and topological statistics.

References

[1] Seungjoo Shin and Jaesik Park. Binary radiance fields. Advances in neural information processing systems, 36:55919–55931, 2023.

[2] Yihang Chen, Qianyi Wu, Mehrtash Harandi, and Jianfei Cai. How far can we compress instant-ngp-based nerf? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20321–20330, 2024.

Comments on the paper

4 comments

We have updated the manuscript and highlighted the revisions in blue. The main text now includes an updated ablation result and a discussion of practical applications and limitations in the Conclusion. Several supplementary experiments are reported in the appendix, with their titles also highlighted in blue. We offer the catalog as follows:

1. Main text: Ablation. We highlight the revised 8-bit residual-coding row at matched geometric accuracy.

2. Main text: Conclusion. We discuss the current costs, offline applications, and future efficiency improvements.

3. Appendix A.4: Sample Strategy. We clarify sampling at each fine hash level and explain how unsampled embeddings participate in training.

4. Appendix A.5: Decoder Side Information. We explain which statistics and search parameters are stored and how the decoder reconstructs feature-pruning masks.

5. Appendix A.6: Baseline Adaptation. We describe how UniTruss and UniMate are adapted to encode large lattices, including their geometry and topology bitstreams.

6. Appendix B: Preprocessing of Regular Grid Points. We clarify how PCA aligns the regular grid and how Hilbert ordering preserves spatial locality.

7. Appendix D.4: Dataset Statistics. We add model statistics, geometric and topological variability measurements, and node-degree histograms.

8. Appendix E.3: Topology Network Ablation. We add comparisons with neighborhood prediction followed by correction-edge coding and with direct coding of the original topology.

9. Appendix F: Limitations. We discuss the GPU memory cost of training and possible directions for improving memory efficiency.

Dear Reviewers,

The authors have posted their rebuttal. Please read it and reply in the discussion, covering:

Whether each of your questions and concerns has been addressed. Whether your rating and recommendation change, and why. Any remaining concern that must be settled before the meta-review. Best.

Dear Reviewers,

Thank you for your time and thoughtful feedback. As the discussion period closes tomorrow, we hope our responses have addressed your concerns. If any questions remain, we would be grateful for the opportunity to clarify them before the discussion ends.

Thank you again for reviewing our paper.

Dear Reviewers,

Thank you for your thoughtful feedback. We have posted a response to your latest concerns and would be grateful if you could review it.

As the discussion period draws to a close, may we ask whether our response has addressed all or most of your concerns? If you find our clarifications satisfactory, we would appreciate your considering an update to your rating. We sincerely look forward to your feedback.

Thank you again for reviewing our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.