acceptodds
Under review as a conference paper at ICLR 2027

HHC-GS: Hyperbolic Hierarchical Context for Compact Multi-Scale Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) enables high-quality rendering with explicit scene representations, but incurs substantial storage costs in large-scale scenes. Existing multiscale methods reduce rendering overhead by selectively activating anchors at different levels of detail, while retaining independently stored high-dimensional features for individual anchors. To reduce this redundancy, we propose HHC-GS, a compact multiscale Gaussian mapping method based on hyperbolic hierarchical context. HHC-GS uses hyperbolic relations to modulate ancestor-conditioned feature prediction, allowing each non-root anchor to store only low-dimensional structural variables and a bounded residual. For local map loading, feature-generation dependencies are restricted to ancestor paths, and dependency closure ensures consistent feature generation between local subsets and the complete map. We further jointly optimize multiple hierarchy prefixes with frozen-teacher supervision to improve reconstruction under different representation budgets while preserving full-map quality. Experiments on MatrixCity, BungeeNeRF, and UrbanScene3D show that HHC-GS maintains competitive reconstruction quality across representation budgets while substantially reducing storage. Compared with Octree-GS, HHC-GS achieves comparable or better reconstruction quality while reducing model size by more than 85% on all three datasets.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D Gaussian Splatting, Hyperbolic Hierarchical Context, Compact Scene Representation, Level-of-Detail Rendering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hE79OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HHC-GS, a compact multiscale Gaussian Splatting framework for large-scale scene reconstruction. It uses hyperbolic hierarchical context to predict anchor features, reducing the need to store full feature vectors. The method also introduces ancestor-closed loading for consistent local reconstruction and multi-prefix training to support different storage budgets. Experiments on three large-scale datasets show over 85% storage reduction compared with Octree-GS while maintaining competitive rendering quality.

Critical strengths
  1. The paper addresses feature redundancy in hierarchical Gaussian representations. The proposed ancestor-conditioned feature prediction provides an effective way to reduce storage.
  2. The experimental results are promising. HHC-GS reduces model size compared with Octree-GS while maintaining competitive rendering quality.
  3. The evaluation is relatively comprehensive. It covers reconstruction quality, storage efficiency, runtime performance, local map loading, and ablation studies.
Critical weaknesses
  1. Unclear source of the overall storage reduction. The ablation study demonstrates the effectiveness of compact feature prediction. However, the Independent Feature variant is already much smaller than Octree-GS (167.3 MB vs. 623.6 MB). The authors should clarify the source of this difference and provide a more detailed storage breakdown.
  2. Limited analysis of hyperbolic modeling. The comparison with Euclidean Context shows promising improvements. However, the contributions of hyperbolic distances and hierarchical relation regularization are not separately evaluated. An additional ablation would help clarify the role of each component.
  3. Missing comparisons with closely related methods. The paper does not compare against ContextGS or HiCo-GS, despite discussing them in the related work. These comparisons are important for evaluating the effectiveness of the proposed hierarchical context modeling.
Questions for authors

My main concerns have been outlined in the Weaknesses section. I would appreciate clarification on these points during the rebuttal.

Reviewer_VkBSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a compact hierarchical Gaussian representation. Hyperbolic distances modulate ancestor aggregation, bounded residuals supply local capacity, and ancestor-closed loading preserves feature dependencies for local rendering. Joint prefix training with a frozen teacher supports different representation budgets. Experiments cover MatrixCity, BungeeNeRF, and UrbanScene3D.

Critical strengths
  1. Ancestor paths define local-loading dependencies, and the paper correctly distinguishes feature consistency from image equality.

  2. Table 1 shows substantial storage savings; Table 4 includes low-dimensional, parent-only, and Euclidean controls that support feasibility.

Critical weaknesses
  1. The source of the storage savings and the quality–storage trade-off need clarification. On MatrixCity, Octree-GS occupies 623.6 MB in Table 1, compared with 167.3 MB for Independent Feature in Table 4 and 82.5 MB for the full method. Relative to Independent Feature, the proposed method roughly halves storage while reducing PSNR from 23.11 to 22.78 dB. Please explain the substantial size difference between the two baselines. Moreover, the second row of Figure 3 suggests that Octree-GS better preserves fine details, while at a larger storage cost. These results insure a clearer assessment of whether the storage savings justify the loss in reconstruction fidelity, considering both PSNR and visual detail.

  2. The geometry control is underspecified. The tree is fixed, while capacity, regularization, and optimization constraints in the Euclidean control are unclear. The entire gain cannot yet be attributed to hyperbolic geometry.

Questions for authors
  1. Can you reconcile Tables 1 and 4 with a storage breakdown using Equation (8): anchor/root counts, feature dimensions, non-feature attributes, precision, indices, and networks? In particular, what accounts for the difference between the 623.6 MB Octree-GS model and the 167.3 MB Independent Feature control? Are topology and training budgets matched?

  2. What exactly replaces hyperbolic modeling in Euclidean Context? Please specify its distance, structural regularizer, constraints, parameter count, and training settings. Does the reported advantage persist under a matched formulation?

  3. Please provide the block construction, full relation-loss settings, prefix sampling distribution, and level-specific filters. Can the authors clarify which are fixed across all datasets and controls?

Reviewer_4GKaOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

It recursively generates anchor features using known octree ancestor relationships, with hyperbolic distances determining the contributions of different ancestors. By storing low-dimensional structural variables and bounded residuals instead of full-dimensional features for non-root anchors, the method reduces storage redundancy.

Critical strengths
  1. The paper clearly presents its motivation and overall pipeline, focusing on reducing the redundancy of high-dimensional anchor features.

  2. The method achieves substantial storage reduction while maintaining competitive reconstruction quality across multiple datasets.

  3. The ablation study is targeted in nature. Table 4 goes beyond simply removing entire modules; it also compares results for "Independent Feature," "Low-dimensional Only," "Parent Only," and "Euclidean Context," as well as the effects of omitting the "Teacher" and "Multi-prefix" components.

Critical weaknesses

1.The independent contribution of hyperbolic modeling could be further clarified. Questions: Although the Euclidean Context ablation demonstrates improvements from hyperbolic modeling, could the authors further disentangle the contributions of hyperbolic distance-based weighting and the hierarchical relation regularizer?

2.In particular, since the ancestor weight is trivially one for level-1 anchors, how do the benefits vary across hierarchy levels?

3.Equation (6) computes hyperbolic distances using the inverse hyperbolic cosine function, whose derivative is singular when its argument approaches one. In finite-precision arithmetic, the direct evaluation of this expression may suffer from cancellation errors and potentially produce non-finite gradients when two embeddings are identical or nearly identical.

Since these distances are used in both ancestor weighting (Eq. 4) and the hierarchical relation loss (Eq. 7), could the authors clarify how numerical stability is ensured during training?

Questions for authors

Please see the weakness.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your time and effort. As the discussion deadline approaches, could you please respond to the authors’ rebuttal and share whether it addresses your concerns? Your continued participation in the subsequent discussions would also be greatly appreciated.

Best regards,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.