acceptodds
Under review as a conference paper at ICLR 2027

SAS-LoD: Selective Attribute Sharing for Efficient Level-of-Detail Gaussian Splatting

Abstract

High-fidelity view synthesis using 3D Gaussian Splatting (3DGS) delivers exceptional visual realism with interactive performance. However, rendering large-scale scenes with millions of Gaussian primitives incurs substantial memory usage and computational overhead, often leading to out-of-memory failures and increased rendering latency on resource-constrained devices. Level-of-detail (LoD) techniques alleviate these bottlenecks, yet existing designs face a trade-off: optimizing separate Gaussian sets inflates the resident memory footprint, whereas sharing a single set causes noticeable quality degradation. In this paper, we overcome this trade-off by observing that scale and opacity, accounting for only 4 of 59 parameters, primarily adapt Gaussian footprints across LoD granularities, while all remaining attributes can be shared without quality loss. Building on these insights, we present an efficient 3DGS framework with Selective Attribute Sharing LoD (SAS-LoD). SAS-LoD maintains a single base Gaussian set with compact level-wise residual offsets for scale and opacity, enabling high-fidelity per-level adaptation with minimal resident memory overhead. Coupled with codebook compression and on-demand decoding, SAS-LoD substantially reduces peak memory footprint and improves rendering throughput with high visual fidelity, enabling real-time deployment even on mobile devices.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Level-of-Detail, Novel View Synthesis

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_ALsrOpenReview ↗
Rating
3(+1)
Confidence
4(+1)
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAS-LoD, which shares Gaussian positions, rotations, and spherical harmonic coefficients across levels while introducing level-specific scale and opacity residuals. Codebook compression and on-demand decoding further reduce resident memory. Experiments on six indoor and outdoor scenes evaluate rendering quality, throughput, and memory.

Critical strengths
  1. A clear representation design. Scale and opacity adaptation plausibly compensates for coverage changes after pruning while avoiding duplication of all attributes.
  2. Controlled evidence for selective sharing. For three levels, Table 7 reduces resident memory from 0.86 to 0.49 GB relative to independent LoD, with only a 0.09 dB PSNR decrease. It recovers much of the quality lost with fully shared attributes.
  3. Benefits persist without quantization. The uncompressed SAS-LoD rows in Table 10 outperform the reported LODGE implementation in Tables 8–9 on throughput, memory, and reconstruction metrics across all six scenes. Resident and working memory are reported separately.
Critical weaknesses
  1. The generality of the central finding is limited. The attribute analysis uses a specific hierarchy constructed by pruning and fine-tuning an optimized representation. Preserving positions, rotations, and appearance while repairing coverage is intuitive under this setup. It remains unclear whether four adaptation parameters remain adequate under stronger simplification or different geometry and appearance. The evidence supports this configuration more strongly than a general principle of Gaussian LoD.
  2. Large-scale application benefits are not fully demonstrated. The outdoor evaluation uses one chunk from each of only two scenes, with out-of-core streaming disabled. It does not measure complete-scene loading, caching, navigation, or peak memory. Uncompressed SAS improves outdoor throughput over LODGE by approximately 5–8%, while both already run at high frame rates; total memory falls by approximately 31–35%. Whether this enables materially larger scenes or fewer loading stalls remains unestablished.
  3. Comparable compression baselines are incomplete. Compression contributes substantially to resident-memory reduction in Table 1. Figure 7(d) does include quantized independent-LoD memory, so this control is not entirely absent. However, the paper lacks a complete quality–throughput–memory comparison with competing LoD methods under comparable compression and on-demand decoding. Such results would isolate the representation’s advantage more convincingly.
Questions for authors

Q1. Coverage of large-scale scenes.
Why is the outdoor evaluation restricted to a single chunk from each of Campus and SmallCity, and how were these chunks selected? Could the authors evaluate all chunks under the same protocol and report both per-chunk results and scene-level aggregates for rendering quality, throughput, and memory? This would help establish whether the reported benefits generalize across the complete scenes. Please distinguish evaluating all chunks individually from rendering the complete scene with streaming enabled.

Q2. Practical applications and significance.
Which practical applications would materially benefit from the proposed quality–speed–memory trade-off? In the outdoor experiments, uncompressed SAS-LoD improves throughput over LODGE by approximately 5–8% and reduces total memory by approximately 31–35%, while both methods already achieve high frame rates. Could the authors demonstrate a concrete setting where these savings enable a previously infeasible workload—for example, rendering a larger scene within a fixed GPU-memory budget—or substantially improve deployment performance? Such evidence would clarify the practical significance of the method beyond relative metric improvements.

Q3. Comparable compression baselines.
A substantial portion of the resident-memory savings comes from compression. Have competing LoD representations been evaluated with comparable quantization and selective decoding? Figure 7(d) provides a quantized independent-LoD memory comparison, but could the authors additionally report rendering quality, throughput, and resident/working memory under matched compression settings? This would help separate the benefits of selective attribute sharing from those of the compression pipeline.

Q4. Novelty and contribution attribution.
Is the main contribution of SAS-LoD the empirical finding about which attributes require adaptation, the compact scale/opacity residual representation, or the selective decoding strategy? The existing attribute ablations are useful; could the authors clarify what is technically new and which controlled comparisons isolate its benefit from the underlying hierarchy construction and compression choices?

9 comments
AuthorsOpenReview ↗

Response to Reviewer ALsr (1/5)

We sincerely appreciate your insightful comments. We have tried to respond to all these valuable concerns as below.


C1. Generality of SAS-LoD

1. Selective attribute sharing for a different LoD

While our attribute analysis in the paper is conducted using a pruning-based hierarchy construction pipeline, we observe that the underlying insight can also extend to other hierarchy construction pipelines. To verify this, we apply our selective attribute sharing strategy to H3DGS, an independent LoD baseline, which merges Gaussians bottom-up into a tree.

As shown in Table R12, integrating selective attribute sharing into H3DGS achieves rendering quality comparable to its original independent representation (25.24 dB vs. 25.30 dB) while reducing resident memory by 22.1%. In contrast, sharing all attributes across levels in H3DGS results in noticeable quality degradation (25.09 dB). This result indicates that our key finding is not restricted to a specific hierarchy construction pipeline and can generalize to other LoD architectures.

Table R12. Ablation study on LoD design with H3DGS (Campus).

MethodMem (GB)Quality
#GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
H3DGS (Independent LoD)382K1030.950.441.3925.300.8110.337
H3DGS (Shared LoD)382K1030.720.441.1625.090.8080.342
H3DGS (SAS-LoD)382K1030.740.441.1825.240.8100.338

2. Stronger simplification

To investigate the effectiveness of SAS-LoD under stronger simplification, we evaluate rendering quality using higher masking thresholds ϵ\epsilon for LoD construction in Table R13. The results indicate that adapting only scale and opacity remains adequate even under stronger simplification. As the masking threshold increases, all three LoD designs contain fewer Gaussians at coarser levels (ll>0), yielding progressively simpler representations. Across all masking thresholds, SAS-LoD maintains rendering quality comparable to independent LoD. In contrast, shared LoD consistently exhibits a noticeable quality gap compared to independent LoD, demonstrating that sharing all attributes fails to compensate for stronger simplification.

Table R13. Ablation study on the masking threshold for LoD construction (Small City).

#G (L1+L2)PSNR ↑SSIM ↑LPIPS ↓
IndepSharedSASIndepSharedSASIndepSharedSASIndepSharedSAS
ϵ=0.02 (default)1562K1537K1526K27.1726.9327.070.8150.8120.8140.2730.2780.275
ϵ=0.05938K943K934K27.0626.8727.030.8140.8100.8130.2760.2810.277
ϵ=0.10520K518K523K26.9326.7626.920.8110.8070.8100.2820.2870.283
ϵ=0.20224K222K225K26.6226.4226.590.8030.7990.8030.2960.3010.296
ϵ=0.30112K110K112K26.1625.9726.170.7930.7890.7930.3130.3190.313

3. Different geometry and appearance

We would like to clarify that our main paper demonstrates the effectiveness of SAS-LoD across diverse geometries and appearances in both outdoor (Hierarchical-3DGS) and indoor (Zip-NeRF) datasets. The outdoor scenes capture large-scale structures, such as buildings, vehicles, and trees along streetscapes, while the indoor scenes encompass multi-room environments (e.g., kitchens, living rooms, and bedrooms) with complex layouts and a wide variety of daily objects.

Furthermore, to complement the outdoor-only ablation study in our main paper (Fig. 7 and Table 9 in the manuscript), we provide an additional ablation study on LoD design using the indoor Zip-NeRF dataset in Table R14. SAS-LoD achieves rendering quality comparable to independent LoD (26.52 dB vs. 26.58 dB), whereas shared LoD incurs a consistent quality degradation (26.37 dB). This consistent result across indoor and outdoor scenes indicates that adapting only scale and opacity serves as a general adaptation strategy across diverse geometries and appearances.

Table R14. Ablation study on LoD designs using the Zip-NeRF dataset.

MethodMem (GB)Quality
#GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
Independent LoD340K2031.060.501.5626.580.8390.319
Shared LoD339K2130.560.501.0626.370.8370.321
SAS-LoD328K2130.590.491.0826.520.8380.320
AuthorsOpenReview ↗

Response to Reviewer ALsr (2/5)

C2. Single-chunk evaluation

We kindly clarify that evaluating a single chunk for each outdoor scene adheres to the standard evaluation protocol of the Hierarchical-3DGS dataset. These chunks are the official evaluation benchmarks designated by the original dataset authors and were not selected by us. Subsequent large-scale 3DGS approaches, including Octree-GS, FLoD, and LODGE, have also adopted this evaluation setup. We therefore follow this protocol to ensure a fair and reproducible comparison with prior work.


C3. Full-chunk evaluation (Large-scale scenes)

We appreciate this insightful suggestion. To verify that our performance gains generalize across complete large-scale scenes, we evaluate all individual chunks of Campus (21 chunks, 22.0K images) and Small City (4 chunks, 5.8K images) under the same per-chunk evaluation protocol. The aggregated results show that the benefits of SAS-LoD are not limited to single benchmark chunks, but consistently generalize across all chunks of large-scale scenes.

We use every 50th image for testing, and all methods are trained and evaluated using this split. We report scene-wise aggregated results in Tables R15 & R16 while providing per-chunk results in the revised manuscript (Sec. C.8) due to space limitations. Quality and FPS are averaged over all test views of all chunks, while memory is reported as the average and maximum over chunks.

Note that this experiment evaluates every chunk individually. Rendering the complete scene, both in-core and with chunk streaming, is evaluated separately in C4.

1. Comparison with baselines across all chunks of each scene

The scene-level aggregate results in Tables R15 & R16 demonstrate that SAS-LoD achieves the highest average rendering throughput and the lowest memory footprint across all chunks, while maintaining rendering quality comparable to the best-performing baseline (LODGE). These aggregated results confirm that the benefits of our method extend across complete large-scale scenes, consistent with our single-chunk evaluations.

Table R15. Quantitative results on all chunks of Campus scene.

MethodMem (GB, Avg / Max)Quality
#GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS1055K1190.65 / 1.030.68 / 0.961.34 / 1.9924.770.7540.378
H3DGS589K1261.38 / 2.180.51 / 0.661.89 / 2.8324.720.7500.384
Octree-GS288K1110.55 / 0.780.53 / 1.021.08 / 1.5824.600.7370.407
FLoD230K1421.00 / 1.280.40 / 0.471.40 / 1.6724.830.7310.416
CLOD-3DGS394K2090.65 / 1.030.35 / 0.461.00 / 1.4924.490.7470.390
LODGE407K1751.08 / 1.700.46 / 0.631.54 / 2.2725.120.7580.372
CLoD-GS424K1310.49 / 0.600.65 / 0.811.14 / 1.3225.090.7540.383
Ours w/o quant.328K2080.68 / 1.070.33 / 0.431.01 / 1.4625.150.7580.372
Ours328K2100.08 / 0.120.33 / 0.430.41 / 0.5425.090.7560.376
Ours (single chunk)331K1860.060.360.4225.940.8170.327

Table R16. Quantitative results on all chunks of Small City.

MethodMem (GB, Avg / Max)Quality
#GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS1026K1050.76 / 0.780.59 / 0.641.35 / 1.4224.550.7680.302
H3DGS409K1491.53 / 1.570.32 / 0.331.85 / 1.8924.400.7610.315
Octree-GS298K1540.69 / 0.900.43 / 0.641.12 / 1.5424.420.7630.309
FLoD269K1970.99 / 1.040.27 / 0.281.26 / 1.3224.290.7450.344
CLOD-3DGS418K1930.76 / 0.780.26 / 0.271.02 / 1.0524.400.7670.306
LODGE291K2851.14 / 1.180.27 / 0.291.41 / 1.4724.800.7720.300
CLoD-GS400K1800.54 / 0.570.37 / 0.390.91 / 0.9624.690.7650.310
Ours w/o quant.253K3350.78 / 0.800.19 / 0.210.97 / 1.0024.810.7720.298
Ours253K3410.09 / 0.090.19 / 0.210.28 / 0.3024.790.7700.302
Ours (single chunk)254K3270.050.200.2626.970.8120.278

(Continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer ALsr (3/5)

C3. Full-chunk evaluation (Large-scale scenes) (continued)

2. Ablation study on LoD design across all chunks of each scene

Moreover, Table R17 presents an ablation study on LoD design conducted across all chunks of Campus and Small City. The results demonstrate that the effectiveness of our selective attribute sharing design also generalizes beyond a single benchmark chunk to complete scenes. Consistent with the single-chunk evaluation, SAS-LoD achieves rendering quality comparable to independent LoD while requiring about 70% of its resident memory. Compared with shared LoD, SAS-LoD achieves superior rendering quality at a similar resident memory cost.

Table R17. Ablation study on LoD designs across all chunks of Campus and Small City.

MethodMem (GB, Avg / Max)Quality
#GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
Campus
Independent LoD341K1981.02 / 1.610.34 / 0.431.36 / 2.0225.330.7590.371
Shared LoD339K2030.65 / 1.030.34 / 0.430.99 / 1.4324.980.7550.375
SAS-LoD328K2080.68 / 1.070.33 / 0.431.01 / 1.4625.150.7580.372
Small City
Independent LoD263K3151.08 / 1.120.20 / 0.211.28 / 1.3224.890.7740.297
Shared LoD262K3270.76 / 0.780.20 / 0.210.96 / 0.9924.680.7690.301
SAS-LoD253K3350.78 / 0.800.19 / 0.210.97 / 1.0024.810.7720.298
Reviewer_kRupOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the memory-quality trade-off in level-of-detail (LoD) rendering for large-scale 3D Gaussian Splatting. Independent LoD representations can adapt every Gaussian attribute at every level but duplicate most parameters; fully shared representations avoid duplication but lose quality when coarse levels simply reuse fine-level attributes. The paper empirically argues that, among the 59 scalar parameters of a degree-3 SH Gaussian, only the three scale parameters and one opacity parameter require substantial level-specific adaptation. SAS-LoD therefore stores one shared base Gaussian set and level-specific residuals only for scale and opacity, while progressively masking Gaussians to construct nested levels. Residual vector quantization and on-demand decoding further reduce resident memory. Experiments on Hierarchical-3DGS and Zip-NeRF scenes report approximately 1.9×1.9\times faster rendering and 2.7×2.7\times lower total memory than the uncompressed 3DGS baseline, with similar image quality, as well as real-time rendering on an iPhone 14 Pro.

Critical strengths
  1. A simple, consequential decomposition of shared versus level-specific attributes. The paper does not merely compress every level uniformly. It asks which attributes actually need LoD-specific freedom and finds that scale and opacity account for most of the quality recovery. This leads to a compact parameterization with only four residual scalars per active Gaussian instead of duplicating all 59 scalars. The ablation in Table 7 supports the intended trade-off: at three LoD levels, SAS-LoD is close to independent LoD quality while keeping resident memory near the shared representation.
  2. Strong end-to-end deployment evidence. The paper measures resident memory, working memory, total memory, FPS, rendered Gaussian count, and image quality rather than reporting storage alone. Table 1 shows that compression and LoD address complementary bottlenecks, while Table 5 demonstrates real-time rendering on an iPhone 14 Pro where the compared baselines fail under the same web renderer. This is unusually concrete evidence for practical relevance.
Critical weaknesses
  1. The efficiency evaluation does not fully characterize preprocessing and temporal behavior. SAS-LoD requires progressive masking/fine-tuning, quantization, and per-frame decoding, but the paper emphasizes runtime memory and FPS. Additional accounting of hierarchy-construction time, compression time would help users understand the overall performance of the method.
Questions for authors

See weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer kRup (1/3)

We sincerely appreciate your insightful comments. We have tried to respond to all these valuable concerns as below.


C1. Analysis of our representation construction time

As the reviewer requested, Table R10 reports the runtime of each stage for constructing our representation: LoD construction and compression. Constructing our compressed LoD representation takes 14.1 min on average, which accounts for approximately 30% of the total pipeline time, including the base 3DGS training (31.6 min on average). Importantly, this one-time construction overhead translates into substantial runtime advantages in memory footprint (2.7×) and rendering speed (1.9×) on the Hierarchical-3DGS dataset.

Table R10. Evaluation on the representation construction time (min).

SceneLoD constructionCompressionTotal
MaskingFine-tuningClusteringFine-tuning
Campus3.39.50.72.215.7
Small City3.86.30.71.712.5
AuthorsOpenReview ↗

Response to Reviewer kRup (2/3)

C2. Comparison of representation construction time with baselines

We further compare representation construction with baseline methods in Table R11. For post-training methods, we report the additional time required after the base 3DGS training.

The runtime comparison indicates that SAS-LoD requires a low computational budget for representation construction compared to existing baselines. Among post-training methods, SAS-LoD without compression achieves an efficient runtime comparable to LODGE (11.5 min vs. 11.3 min), and even with compression, it remains faster than all other post-training methods except LODGE.

Table R11. Runtime comparison (min).

MethodCampusSmall CityAverage
3DGS37.725.531.6
From-scratch
Octree-GS35.937.436.7
CLoD-GS68.446.457.4
FLoD115.372.293.8
Post-training
H3DGS19.612.015.8
CLOD-3DGS25.414.620.0
LODGE12.99.711.3
Ours w/o quant.12.810.111.5
Ours15.712.514.1
AuthorsOpenReview ↗

Response to Reviewer kRup (3/3)

C3. Per-frame decoding overhead

Our on-demand decoding retrieves the attributes of only the Gaussians selected for the current view, which requires just a few codebook lookups per Gaussian. Its per-frame overhead is therefore negligible. As shown in Table R20 in our response to Reviewer ALsr, despite this additional decoding, our compressed representation renders slightly faster than the uncompressed one (257 FPS vs. 252 FPS), as the compact codebook indices reduce the memory bandwidth required to fetch Gaussian attributes. At the same time, it reduces resident memory by 9.1×. Note that all frame rates reported for our method include both LoD selection and decoding.

Thanks for the extra analysis regarding efficiency analysis. My concern has been well resolved, and I think supplementing these materials to the paper would help provide a comprehensive understanding of the method. Thus, I would maintain my weak acceptance opinion.

Reviewer_X5VDOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies level-of-detail rendering for large-scale 3D Gaussian Splatting. Independent per-level Gaussian sets give high quality but redundant storage, while sharing one set saves memory at the cost of quality. Through a controlled attribute-sharing study, the authors find that per-level adaptation is concentrated in scale and opacity, whereas position, rotation, and spherical-harmonics coefficients can be shared across levels with little loss. SAS-LoD builds directly on this finding: a single base Gaussian set with nested active subsets, compact level-wise residuals for scale and opacity only, codebook compression of all attributes, and on-demand decoding of the Gaussians selected for the current view. Headline results — quality close to independent hierarchies, lower total memory, faster rendering, and real-time rendering on a phone where baselines fail — are reported with compression enabled, while the controlled comparison of LoD designs is reported without compression. Experiments use two outdoor chunks and several indoor scenes, with ablations and a mobile deployment study.

Critical strengths
  1. The design is driven by an empirical finding rather than a heuristic, and the resulting representation is simple and deployable: one base set plus a few values per Gaussian per extra level, with contiguous prefix storage that composes naturally with quantization and view-dependent decoding. The uncompressed design comparison is informative and the paper reports the unfavorable directions honestly.
  2. Validation covers the axes that matter for deployment: quality, resident memory, working memory, and rendering speed, with per-scene tables and per-attribute memory accounting that make the efficiency claims checkable, plus a mobile deployment study that addresses practical relevance.
Critical weaknesses
  1. The central trade-off claim is not established in the deployed configuration. All headline results use codebook compression, but the only controlled comparison of LoD designs is uncompressed. Under the same compression, resident memory becomes a small share of the total, so the total-memory advantage over an equally compressed independent hierarchy shrinks to a small margin and essentially vanishes against an equally compressed shared hierarchy, while the uncompressed quality ordering may not survive quantization. The paper never compares the three designs under identical compression, which is the configuration it actually advocates. In addition, the uncompressed comparison may not isolate the sharing variable itself, since the active sets of the compared designs appear to differ.
  2. The key empirical finding is narrower than presented, and selective adaptation is not new per se. The sharing study is post-hoc — it swaps attributes without re-optimization — and post-hoc sensitivity does not establish necessity under joint optimization; the sequential-sharing result is order-dependent; and with degree-3 harmonics most of the parameter saving comes from sharing harmonic coefficients, so the benefit shrinks at lower degrees. Moreover, the claim that prior shared hierarchies adapt either every attribute or none is contradicted by work the paper itself cites that adapts opacity only, and by recent layered and budget-flexible splatting methods that adapt proper subsets of attributes; none of these are discussed. The novelty is therefore the specific attribute choice plus integration, and the headline speedup comes from the two outdoor chunks while gains on the indoor scenes are much smaller.
Questions for authors
  1. Can you compare independent, shared, and selective sharing under identical compression and on-demand decoding, reporting quality and total memory? The quantized resident-memory figures already exist; the corresponding quality does not.
  2. In the uncompressed design comparison, were the active index sets matched across designs? If masks are recomputed per design, the comparison mixes attribute freedom with different retained sets.
  3. Does the sharing finding survive joint optimization, for instance a jointly trained opacity-only or scale-only variant at matched memory, and does the sequential-sharing ranking hold under a different schedule or at lower harmonic degrees?
  4. How do you reconcile the all-or-nothing characterization of prior shared hierarchies with opacity-only adaptation in the work you cite, and with layered or budget-flexible splatting methods that adapt subsets of attributes?
  5. For the mobile study: which renderer version and harmonic degree were used, and at what precision and resolution? Does the reported frame rate include selection and decoding, why exactly do the baselines fail, is the natural control of plain 3DGS with the same compression available, and is the listed device memory correct?
  6. The abstract says the remaining attributes can be shared without quality loss while the contribution list says with modest loss; please align these, and clarify how the headline speedup aggregates across scenes with very different gains.
9 comments
AuthorsOpenReview ↗

Response to Reviewer X5VD

We sincerely appreciate your insightful comments on the summary and strengths, which accurately capture the core ideas of our paper.

However, we noticed that the weaknesses and questions seem to refer to a different submission. Specifically, they mention components that are unrelated to our work, such as GPT-2 configurations and text benchmarks.

To ensure that we can properly address your concerns during the rebuttal, we would be grateful if you could review this and update the weaknesses and questions with comments relevant to our paper.

AuthorsOpenReview ↗

Response to Reviewer X5VD (1/7)

We sincerely appreciate your insightful comments. We have tried to respond to all these valuable concerns as below.


C1. Comparison of LoD designs under compression

Please refer to C1 in the global response for the full comparison.

As shown in Tables R1 and R2, the quality ranking is preserved under compression. SAS-LoD still achieves rendering quality comparable to independent LoD while requiring total memory comparable to shared LoD.

Total memory advantage under compression

We agree that the total memory advantage over independent LoD appears modest under compression. At the full-scene scale, however, our LoD design can still yield substantial total memory savings even with compression. Table R3 compares the memory footprint of LoD designs on Campus across two settings: (a) a single chunk with compression and (b) the complete scene (all 21 chunks, 22.0K images) with compression.

Under compression, SAS-LoD saves only 0.041 GB of resident memory over independent LoD within a single chunk (a).When rendering the complete large-scale scene (b), the saving grows by 18.5× (to 0.759 GB), as resident memory scales with the scene size. Since working memory is bounded by the visible region in the current view rather than the total scene size, this resident memory reduction translates directly into a comparable reduction in total memory. These results demonstrate that, for complete large-scale scenes, SAS-LoD maintains a substantial memory advantage even under compression.

Table R3. Resident memory (GB) comparison of LoD designs under compression for single-chunk and full-scene rendering on Campus.

Method(a) Single chunk(b) Full scene
Independent LoD0.0982.380
Shared LoD0.050 (-0.048)1.525 (-0.855)
SAS-LoD0.057 (-0.041)1.621 (-0.759)
AuthorsOpenReview ↗

Response to Reviewer X5VD (2/7)

C2. Matched active sets across LoD designs

We appreciate this insightful comment. In the original comparison, the active masks were computed independently for each design according to the masking threshold. Thus, to isolate the effect of attribute sharing, we conduct an ablation study on the LoD design with matched active sets in Table R4. Note that '#Active G' represents the total number of active Gaussians across all levels.

Specifically, we first construct SAS-LoD following our original training procedure to extract the mask at each step, and then apply these fixed per-step masks to all compared LoD designs. This ensures that every design employs the exact same set of Gaussians at each level and differs only in which attributes are adapted across levels. All designs are then jointly fine-tuned under the same protocol.

The results show that our finding holds consistently under matched active sets, as SAS-LoD achieves quality comparable to independent LoD (26.60 dB vs. 26.69 dB), whereas shared LoD still shows a noticeable degradation (26.49 dB).

Table R4. Ablation study on LoD design with matched active sets. Our default setting (L=3L=3) is highlighted in bold.

MethodMem (GB)Quality
#Param#Active GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
L = 2
Independent LoD198.3M3.36M1930.740.341.0826.460.8140.301
Shared LoD123.5M3.36M1950.460.330.8026.280.8120.304
SAS-LoD128.5M3.36M1950.480.340.8226.390.8140.302
L = 3
Independent LoD226.5M3.84M2420.850.281.1226.690.8170.298
Shared LoD123.5M3.84M2510.460.280.7426.490.8140.302
SAS-LoD130.5M3.84M2480.490.280.7726.600.8160.299
L = 4
Independent LoD235.1M3.98M2490.880.271.1526.790.8180.296
Shared LoD123.5M3.98M2590.460.270.7426.530.8150.301
SAS-LoD131.0M3.98M2550.490.280.7626.650.8170.298
L = 5
Independent LoD238.0M4.03M2490.890.281.1626.930.8190.294
Shared LoD123.5M4.03M2570.460.280.7426.650.8160.299
SAS-LoD131.2M4.03M2550.490.280.7726.870.8180.296

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response by Authors

Global Response

Dear reviewers,

We thank all reviewers for their thoughtful comments. We have tried to respond to all these valuable concerns. In response to the comments, we have revised the manuscript with additional experimental results. Please note that we have highlighted updates in blue.


C1. Comparison of LoD designs under a compressed setting (X5VD, ALsr)

As Reviewers X5VD and ALsr requested, we report evaluation results under a compressed setting in Tables R1 and R2.

1. Ablation study on LoD design under a compressed setting

Table R1 presents an ablation study on LoD design under the matched compression setting by integrating our compression and on-demand decoding schemes. Even under the matched compression setting, the relative performance in rendering quality, throughput, and memory is consistent with that of the uncompressed setting. Specifically, our method achieves rendering quality comparable to independent LoD while reducing resident memory usage to 60%. Conversely, shared LoD still exhibits a noticeable quality degradation despite its slightly lower resident memory footprint.

Table R1. Ablation study on LoD design under a compressed setting (L=3L=3).

MethodMem (MB)Quality
#GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
Independent LoD300K2439329338626.480.8150.302
Shared LoD298K2515029134126.310.8130.307
SAS-LoD292K2575528934426.460.8140.303

2. Comparison with baselines under a compressed setting

We further evaluate baseline LoD methods under matched compression settings in Table R2. We exclude Octree-GS from this experiment because its anchor-based representation, encoded via MLPs, is structurally incompatible with our Gaussian attribute compression scheme.

Table R2 reports the quantitative results under this matched compression setup, including rendering quality, throughput, and resident and working memory. Applying on-demand decoding keeps rendering speeds nearly unchanged or slightly improved across all methods, with our method still achieving the highest rendering speed.

Among shared LoD representations, CLOD-3DGS achieves the lowest resident memory since it introduces no additional parameters beyond the input Gaussians. However, this parameter sharing comes at the cost of noticeable degradation in rendering quality (25.41 dB vs. 26.06 dB of 3DGS). SAS-LoD follows closely in resident memory, requiring only approximately 10% more resident memory than CLOD-3DGS while preserving high visual fidelity.

Compared to independent LoD baselines, SAS-LoD achieves the highest rendering quality while requiring the lowest resident and working memory. Specifically, all independent LoD representations require at least 100 MB of resident memory, whereas SAS-LoD maintains high rendering quality with only 55 MB.

These results demonstrate that SAS-LoD retains its advantages in rendering efficiency under a matched compression pipeline, indicating that its core benefits result from selective attribute sharing rather than the compression scheme itself.

Table R2. Comparison with baselines under a compressed setting.

MethodMem (MB)Quality
#GFPS ↑Res ↓Work ↓Total ↓PSNR ↑SSIM ↑LPIPS ↓
Single-level
3DGS685K1335046151126.060.8100.307
Shared LoD
CLOD-3DGS335K2205028533525.410.7970.329
CLoD-GS381K1636347754025.890.8050.320
Independent LoD
H3DGS351K13931134966025.790.8060.317
FLoD225K18510233543725.480.7890.349
LODGE315K23510333744026.140.8110.308
SAS-LoD
Ours292K2575528934426.460.8140.303

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.