acceptodds
Under review as a conference paper at ICLR 2027

From Geometry to Semantics: Efficient 3DGS Feature Compression with Geometry-Transferred Low-Rank Adaptation

Abstract

3D Gaussian Splatting (3DGS) enables high-fidelity rendering by representing scenes with explicit Gaussian primitives, but its extension to semantic-aware representations introduces substantial storage overhead due to high-dimensional language-aligned features. Existing compression methods, designed primarily for geometric fidelity, fail to preserve task-relevant semantics under aggressive bitrate constraints, revealing a fundamental mismatch between signal reconstruction and feature geometry preservation. In this work, we revisit semantic compression from a task-oriented perspective and propose a rate–distortion optimized framework for semantic feature compression in 3DGS. Our key insight is that effective compression must preserve the geometry of language-aligned feature spaces rather than raw signal fidelity. To this end, we transfer geometric compression priors to the semantic domain via low-rank adaptation (LoRA), enabling parameter-efficient alignment without modifying the pre-trained backbone. Furthermore, we introduce a geometry-guided attention mechanism that leverages geometric latents as zero-cost side information to enhance semantic consistency during compression. The entire framework is trained end-to-end under a hyperprior-based entropy model, enabling explicit bitrate control. Extensive experiments show that our method achieves near-lossless downstream performance (mIoU 0.812 vs. 0.816 uncompressed) at a significantly reduced bitrate (3.67 bpe), while preserving downstream capabilities such as semantic retrieval and language-guided editing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D gaussian splatting, computer vision, lora, semantics

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_4FinOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

The paper proposes a framework for compressing semantic features attached to 3D Gaussian representations by adapting a pretrained FCGS compressor. Nonlinear projection encoder bridges the semantic and attribute feature spaces, while LoRA updates adapt the frozen feature transforms and hyperprior. An attention mechanism gated module further refines semantic latents before quantization. Experiments report near-lossless segmentation performance on Replica scenes, including 0.8124 mIoU at 3.67 BPE on room_0 versus 0.8164 uncompressed. Additional experiments investigate projection-only adaptation across scenes, semantic retrieval, qualitative editing, and compression of DF-3DGS features.

Critical strengths
  1. The method maintains performance close to the uncompressed reference across multiple operating points on both evaluated Replica scenes. Its substantial gains over FCGS-Semantic-RD provide evidence that the complete adaptation framework is better.

  2. The framework separates source-trained compression components from scene-dependent projection networks. Updating only the projections recovers useful reconstruction and segmentation performance on additional scenes, providing concrete evidence that the learned compression components can be reused.

Critical weaknesses

The evaluation is restricted to DF-3DGS, whose scene-specific compression module is omitted. Comparisons with other relevant approaches are also missing, limiting the evidence that the proposed framework offers a competitive compression–accuracy trade-off beyond this particular baseline. Broader benchmarking under consistent evaluation and storage conventions is needed to substantiate the claimed advantages.

Questions for authors
  1. How much of the improvement is attributable to LoRA itself? Under the same architecture and training conditions, does LoRA achieve comparable rate–semantic performance to full fine-tuning with less adaptation time or lower peak memory usage?

  2. Do the evaluated baselines fully exploit their compression capabilities? In particular, why is DF-3DGS evaluated without its scene-specific compression module, and do the reported advantages persist when this component is included?

7 comments

Thank you for recognizing semantic preservation and cross-scene reuse. We address the complete DF-3DGS baseline and the contribution of LoRA below.

1. Complete DF-3DGS comparison and storage accounting

Our semantic-accuracy advantage is retained when DF-3DGS's scene-specific compression is enabled. The original table in our paper emphasized semantic preservation and reported its higher-mIoU configuration without SC. We now include our reproduction of the complete, unmodified method, evaluated on Replica room0 with the same protocol as the manuscript's main comparison.

Following DF-3DGS's storage convention (Sec. 4.1), its 13.1 MB includes the complete semantic-field PLY (positions, scales, rotations, opacities, and semantic features) plus decoder weights, excluding the separate RGB field. For our coupled representation, we count semantic data and codec weights, excluding geometry/appearance shared with RGB.

MethodSemantic storage (MB)mIoU
DF-3DGS, including scene-specific compression13.100.7651
Ours (λ=10−4\lambda=10^{-4})59.850.8124

Our entry is calculated from the manuscript's storage breakdown: 55.65 MB semantic data + 1.11 MB scene-specific parameters + 3.09 MB shared backbone, counted once. The 18.35 MB geometry/appearance component is excluded. With the backbone already available, per-scene semantic storage is 56.76 MB.

At these operating points, our method achieves 4.73 percentage points higher mIoU and remains close to its uncompressed reference (0.8164). This supports semantic preservation under compression. The table compares complete semantic representations with different storage budgets; controlled comparisons on fixed source features are discussed below. We will add the complete DF-3DGS result to Table 1 and specify the storage components.

Appendix E of our paper further demonstrates applicability to DF-3DGS's 512-D source features: our compression retains 0.764 mIoU versus 0.774 before compression.

2. LoRA versus full fine-tuning under the same architecture

To isolate the adaptation strategy, both variants use the same projection, GeoGuide, entropy modules, and loss, start from the same Stage 1 checkpoint, and receive the same number of training updates. The five operating points below use the same evaluation protocol on Replica room0. BPE denotes estimated semantic rate.

λ\lambdaLoRA BPELoRA mIoUFull FT BPEFull FT mIoU
1e-43.66970.81243.53100.8113
2e-43.55140.81173.43860.8099
4e-43.17690.80793.09170.8058
8e-43.00580.80772.90950.8049
16e-42.94250.80652.86180.7912

These results support competitive rate–semantic performance. For example, LoRA obtains 0.8077 at 3.0058 BPE, versus 0.8058 at 3.0917 BPE for full fine-tuning: 2.78% fewer estimated bits with 0.19 percentage points higher mIoU. At the higher-quality end, LoRA obtains 0.8117 at 3.5514 BPE, versus 0.8113 at 3.5310 BPE. Comparisons use achieved rates, since equal λ\lambda does not imply equal BPE.

A separate controlled implementation profile on an NVIDIA A100 measures adaptation costs for the same architecture:

Adaptation costLoRAFull FT
All trainable parameters276,792997,688
Backbone adaptation parameters50,624771,520
FP32 backbone adaptation weights (MB)0.20253.0861
Mean Stage 2 optimization time (s)103.4796.53
Peak allocated GPU memory (GiB)3.12933.1149

The profile averages an earlier six-rate sweep with 1,500 Stage 2 updates per setting, excluding checkpoint I/O. LoRA/Full FT learning rates are 10−310^{-3}/10−410^{-4}. These costs are profiled separately from the five-point quality table. LoRA reduces adaptation-weight storage, but provides no training-time or peak-memory advantage in the current implementation. It reduces total trainable parameters by 72.26% and backbone adaptation parameters by 93.44%. We will add this comparison and describe LoRA's benefit as compact adaptation with competitive semantic preservation.

3. Task scope and baseline selection

Our primary task is to compress an existing semantic Gaussian representation while preserving its feature utility. We compare SVD and FCGS-Semantic-RD on fixed source features in room0 and office3. FCGS-Semantic-RD is fully fine-tuned with semantic reconstruction and rate losses, providing a learned compression baseline adapted to semantic data. On room0, SVD and FCGS-Semantic-RD obtain 0.762 mIoU at 4.00 BPE and 0.716 at 3.61 BPE, respectively, versus our 0.8124 at 3.67 BPE.

Complete DF-3DGS additionally evaluates a learned compact semantic representation, providing a system-level storage–accuracy comparison. We will clarify these complementary comparison settings and their storage accounting.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than three days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Area_Chair_BsCFOpenReview ↗

Please review the authors' responses and engage in the discussion

Dear Reviewers,

The authors have now responded to your concerns. It is time to review their responses and assess whether they adequately address your concerns or whether any further questions remain.

I would kindly ask you to read the authors' responses and actively engage in the discussion, as there are only a few days remaining.

Thank you for your time and contribution.

Best regards, Your AC

Reviewer_728dOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to solve the compression of semantic features in 3DGS, aiming to reduce the storage overhead while preserving its downstream semantic capabilities. Existing Gaussian parameter compressors suffer from dimensional and distributional mismatch, the key idea of this paper is to adapt a pretrained compressor to preserve semantic features rather than only minimize the reconstruction error, by connecting semantic features to a frozen FCGS backbone through nonlinear projections and LoRA.

Critical strengths

This paper proposes a practical approach to achieve a good trade-off between the additional storage overhead and semantic-modeling capability of Gaussian parameters, the paper is clear and easy to follow, and the results are promising,

Critical weaknesses
  1. Ablation issues: although the GeoGuide ablation indicates improvement, this improvement could also come from more bits, which makes it hard to tell the effectiveness of the method. And the full-finetuning comparison also changes the adapter architecture which does not leave the LoRA alone.
  2. Misleading related works: the mentioned LangSplat uses a low-dimensional autoencoder representation but is grouped with methods storing raw high-dimensional features.
  3. Unclear contribution: I cannot tell the individual contribution to the improvement of the overall results, between the proposed entropy model, the pretrained compressor, and the compact semantic representations themselves
Questions for authors
  1. Provide a GeoGuide ablation conducted at matched rate.
  2. Clarify the adaptation claim, is it about the LoRA specifically or about the whole architecture?
  3. Correct the misleading related work.
  4. Clarify the comparison boundary between compressing fixed source features and learning compact language fields, including the DF-3DGS variant used in the table.
3 comments

Thank you for recognizing the semantic preservation results and the distinctions in our evaluation. We address the attribution of gains and the comparison boundaries below.

1. GeoGuide at comparable achieved rates

The improvement persists at closely comparable achieved rates. We extend the room0 ablation to five operating points. Both variants retain learned projection, LoRA, and entropy modeling, and use the same views, labels, feature decoder, and mIoU evaluation. BPE denotes estimated semantic rate.

λ\lambdaFull model BPEFull model mIoUNo GeoGuide BPENo GeoGuide mIoU
1e-43.66970.81242.98460.7613
2e-43.55140.81172.97380.7586
4e-43.17690.80792.95610.7532
8e-43.00580.80772.92490.7494
16e-42.94250.80652.93530.7451

The full model achieves 0.8065 mIoU at 2.9425 BPE, compared with 0.7532 at 2.9561 BPE without GeoGuide: 5.33 percentage points higher mIoU with 0.46% fewer estimated bits. This measured comparison addresses the concern that the original gain could simply come from spending more bits. We will add these operating points alongside Table 7 in our paper and distinguish this closely comparable-rate result from the original equal-λ\lambda comparison.

2. Scope of the adaptation claim and same-architecture comparison

Our main contribution is the complete architecture for reusing pretrained Gaussian compression transforms on supplied semantic features. LoRA provides compact adaptation within this architecture. The original FCGS-Semantic-RD comparison evaluates the complete framework; we agree that its different adapters do not isolate LoRA.

We therefore compare LoRA with full fine-tuning using the same projection, GeoGuide, entropy modules, and loss, a shared Stage 1 checkpoint, equal training-update counts, and the same evaluation protocol.

λ\lambdaLoRA BPELoRA mIoUFull FT BPEFull FT mIoU
1e-43.66970.81243.53100.8113
2e-43.55140.81173.43860.8099
4e-43.17690.80793.09170.8058
8e-43.00580.80772.90950.8049
16e-42.94250.80652.86180.7912

LoRA obtains 0.8077 at 3.0058 BPE, versus 0.8058 at 3.0917 BPE for full fine-tuning, supporting competitive rate–semantic performance. The backbone adaptation uses 50,624 versus 771,520 trainable parameters, a 93.44% reduction. We will distinguish this parameter-efficiency result from the complete architecture's performance gains, and add the controlled comparison to the adaptation discussion.

3. LangSplat correction and task boundary

Thanks for pointing this out. LangSplat uses a scene-specific autoencoder and stores low-dimensional language features in its Gaussians; decoding the rendered features recovers CLIP embeddings. We will correct Section 2 and remove the blanket description that these methods store raw high-dimensional features at every Gaussian.

Our contribution concerns adapting pretrained compression transforms and optimizing a hyperprior-based semantic rate–distortion objective. We will distinguish this contribution from the established idea of compact semantic representations. Specifically, compressing fixed source features tests preservation of an existing representation, whereas learning a compact language field also changes how the representation is constructed. These settings provide complementary codec-level and system-level comparisons.

4. DF-3DGS variant and complete-method comparison

The DF-3DGS row in Table 1 is explicitly the w/o SC variant, omitting its scene-specific autoencoder. We will clarify this distinction and add our reproduction of complete DF-3DGS with SC enabled, evaluated using the same room0 protocol as the main comparison.

MethodSemantic storage (MB)mIoU
DF-3DGS, including SC13.100.7651
Ours (λ=10−4\lambda=10^{-4})59.850.8124

Following DF-3DGS's accounting, its 13.1 MB includes all attributes of the semantic-field Gaussians and decoder weights, excluding the separate RGB field. Our entry is 55.65 MB semantic data + 1.11 MB scene-specific parameters + 3.09 MB shared backbone, counted once; RGB-shared geometry/appearance is excluded. The table compares complete semantic representations at their reported storage budgets.

Appendix E of our paper addresses a separate question: our compressor applied to DF-3DGS's 512-D source features retains 0.764 mIoU versus 0.774 before compression. We will clearly separate this applicability experiment from the complete-method comparison.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than three days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

The authors’ response addresses my concerns. I maintain my score.

Reviewer_Y4xyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The proposed method shifts the focus from raw signal fidelity to preserving the structural geometry of language-aligned feature spaces. To achieve this efficiently, the authors leverage Low-Rank Adaptation (LoRA) to transfer geometric compression priors to the semantic domain without altering the pretrained backbone. Furthermore, they introduce a geometry-guided attention mechanism that utilizes existing geometric data as zero-cost side information to ensure semantic consistency during the compression process.

Critical strengths
  • The proposed method utilizes geometric latents to guide semantic latent coding, enhancing performance without incurring any additional bitrate overhead.

  • Employing LoRA allows the model to adapt frozen pretrained backbones without retraining heavy feature extractors for each 3D scene.

Critical weaknesses
  • CSGaussian (WACV26) already proposed an RD-optimized compression for both geometry and semantic features. Unfortunately, it was not properly recognized and included for comparison.

TSeng et al., "CSGaussian: Progressive Rate-Distortion Compression and Segmentation for 3D Gaussian Splatting," WACV 2026

  • The necessity of using LoRA is unclear. The encoder and decoder for geometry compression appear to be relatively lightweight and easy to train. I also believe this is also the case with the semantic part. I do not see why LoRA would be needed as compared to full fine tuning. LoRA's benefits in terms of training, compression overhead, and computational cost (KMAC/pixel) are not clear.
Questions for authors
  • Could you provide the full RD curves to evaluate the effectiveness of each module you have added?

  • Could you compare the performance of the proposed framework with CSGaussian?

  • The rational of including w_i to weight semantic features differently needs justification.

  • The writing quality needs improvement. For instance, in the system overview section (Section 3.1: Overview), many symbols mentioned in the text do not appear in Figure 1. Conversely, the figure also contains several symbols that are not explained in the text. Please update

  • Figure 1 and the corresponding descriptive text to make them easier to understand. What is the rendering speed (frames per second) at the decoding stage of this model? Does it still maintain the real-time performance characteristic of Gaussian Splatting?

4 comments

1. Relationship to CSGaussian

We thank the reviewer for pointing out CSGaussian. An earlier version of our work was submitted to NeurIPS in early May 2026, less than four months after CSGaussian’s initial arXiv posting on January 19, 2026. We will cite CSGaussian explicitly, clarify its relationship to our work, and remove the broad claim of being the first semantic rate–distortion compression framework.

The two methods differ in where the semantic representation comes from. CSGaussian learns semantic features under compression constraints within a progressive appearance-and-semantics pipeline. Our method takes an existing language-aligned feature representation as input and asks how much of its semantic information can be retained at a given bitrate by adapting pretrained Gaussian compression transforms. Our technical contribution is the design that enables this reuse: projection maps the supplied features into a representation compatible with the pretrained transforms, LoRA adapts those transforms to the projected features, and attribute-conditioned refinement uses Gaussian attributes to improve semantic reconstruction. These components are jointly optimized under a semantic rate–distortion objective while the pretrained backbone weights remain frozen.

Our framework also supports feed-forward compression across scenes. Once adapted on a source scene, the learned codec can be applied directly to unseen scenes in the same dataset without target-scene parameter updates. We evaluate this zero-shot setting in Table 3. This differs from CSGaussian's pipeline, which learns and compresses scene-specific semantic features through per-scene optimization.

We could not locate an official public implementation of CSGaussian. It reports dataset-level curves for LERF and 3D-OVS without per-scene numerical results and the main text does not specify how results across scenes are aggregated in these curves. The published results therefore provide dataset-level context but do not allow us to establish an aligned per-scene comparison from the reported numbers alone.

2. LoRA versus full fine-tuning: quality and adaptation cost

LoRA is intended to provide compact scene-specific adaptation of the pretrained compressor. We do not claim that full fine-tuning is infeasible. The five-point comparison below is evaluated under the same protocol on Replica room0. Both variants start from the same Stage 1 checkpoint and receive the same number of training updates. The results support comparable semantic preservation, while the separate implementation profile quantifies the reduction in adaptation parameters.

Table R1. LoRA versus full fine-tuning

λ\lambdaLoRA BPELoRA mIoUFull FT BPEFull FT mIoU
1e-43.66970.81243.53100.8113
2e-43.55140.81173.43860.8099
4e-43.17690.80793.09170.8058
8e-43.00580.80772.90950.8049
16e-42.94250.80652.86180.7912

Comparison by achieved rate. A direct comparison of measured operating points is more informative: LoRA achieves 0.8077 mIoU at 3.0058 BPE, whereas full fine-tuning achieves 0.8058 at 3.0917 BPE. Thus, this LoRA point uses 2.78% fewer estimated bits and has 0.19 percentage points higher mIoU. At the higher-quality end, the results are close: 0.8117 at 3.5514 BPE for LoRA versus 0.8113 at 3.5310 BPE for full fine-tuning, a rate difference of approximately 0.58%. These observations support a competitive RD trade-off; they do not establish statistical significance or superiority at every rate.

Adaptation overhead. A separate profile of the complete architecture, with the same projection, GeoGuide, and entropy modules in both variants, gives the following costs. These measurements come from the earlier controlled runs, not from the 5 checkpoints in Table R1.

Table R2. Separate implementation profile on A100

Adaptation costLoRAFull FT
All trainable parameters276,792997,688
Backbone adaptation parameters50,624771,520
FP32 backbone adaptation tensors (MB)0.2024963.086080
Saved model checkpoint (MB; profiled λ = 10⁻⁴)1.1317684.014840
Mean Stage 2 optimization time (s; earlier six-rate sweep)103.4796.53
Peak allocated GPU memory (GiB)3.12933.1149

This implementation reduces total trainable parameters by 72.26% and backbone adaptation parameters by 93.44%. Timing excludes checkpoint I/O; the common Stage 1 costs approximately 23.98 s. We will add the RD comparison and explicit cost accounting, and state the benefit as compact adaptation with competitive semantic preservation. Rendering and neural-operation costs are reported in Section 6 below.

3. Full component RD curves

We provide five evaluated operating points per configuration in Tables R3–R5. The common baseline, Full model w/o GeoGuide, enables learned projection, LoRA updates, and entropy modeling. In Table R4, the other variant additionally disables LoRA updates while retaining Stage 2 optimization of projection and the entropy bottleneck. In Table R5, the other variant instead replaces learned projection with fixed channel truncation. These are comparisons on the explicitly stated background architecture, not three leave-one-component-out variants of the full model.

Table R3. GeoGuide: full model versus full model w/o GeoGuide.

λ\lambdaFull model BPEFull model mIoUFull model w/o GeoGuide BPEFull model w/o GeoGuide mIoU
1e-43.66970.81242.98460.7613
2e-43.55140.81172.97380.7586
4e-43.17690.80792.95610.7532
8e-43.00580.80772.92490.7494
16e-42.94250.80652.93530.7451

Table R4. LoRA updates: GeoGuide is disabled in both variants.

λ\lambdaFull model w/o GeoGuide BPEFull model w/o GeoGuide mIoUNo LoRA updates BPENo LoRA updates mIoU
1e-42.98460.76137.22210.7071
2e-42.97380.75867.18720.7048
4e-42.95610.75326.64350.6972
8e-42.92490.74946.09470.6931
16e-42.93530.74515.88400.6869

Table R5. Projection: GeoGuide is disabled and LoRA is enabled in both variants.

λ\lambdaFull model w/o GeoGuide BPEFull model w/o GeoGuide mIoUFixed truncation BPEFixed truncation mIoU
1e-42.98460.76133.01980.5028
2e-42.97380.75862.91270.5025
4e-42.95610.75322.75190.4982
8e-42.92490.74942.64390.4767
16e-42.93530.74512.47760.4452

GeoGuide. The full model therefore uses 0.46% fewer estimated bits while improving mIoU by 5.33 percentage points. This measured comparison addresses the concern that the original Table 7 improvement could simply result from spending more bits.

LoRA updates. Every supplied LoRA-enabled point has a lower estimated rate and higher mIoU than the supplied frozen-backbone points. This supports adaptation of the pretrained transforms under this architecture and evaluation setting.

Learned projection. The result supports the learned-projection configuration over the tested fixed mapping. The comparison includes the learned projection's warm-up and does not isolate nonlinearity from linear projection, or establish superiority at equal total training cost.

The entropy mechanism remains enabled in these comparisons; these results do not separately isolate its contribution. The weighting ablation is reported separately below.

4. Rationale and ablation for wᵢ

We use wiw_i as a normalized reconstruction weight: it assigns greater MSE penalty to errors in higher-norm source features, complementing the directional preservation encouraged by cosine loss. Feature magnitudes can affect the rendered feature mixture, which motivates this magnitude-sensitive heuristic. Normalization keeps the mean weight approximately one. We do not interpret feature norm as a calibrated measure of semantic importance or confidence.

wi=∥si∥21N∑j=1N∥sj∥2+εw_i = \frac{\left\|s_i\right\|_2}{\frac{1}{N}\sum_{j=1}^{N}\left\|s_j\right\|_2 + \varepsilon}

Here, sis_i is the source semantic feature of Gaussian ii, and NN is the number of Gaussians. In our loss, wiw_i multiplies only the MSE term; it does not weight the cosine, smoothness, or rate terms.

We compare this weighting with uniform weights (wi=1w_i=1) in the complete model, including projection, LoRA, GeoGuide, and entropy modeling. Table R6 reports all five operating points.

Table R6. Feature-norm versus uniform MSE weighting.

Rate weight λ\lambdaWeighted BPEWeighted mIoUUniform BPEUniform mIoU
1e-43.66970.81243.98980.7787
2e-43.55140.81173.96450.7746
4e-43.17690.80793.91670.7724
8e-43.00580.80773.80380.7629
16e-42.94250.80653.55860.7614

Across the five reported lambdalambda settings, norm weighting achieves both lower BPE and higher mIoU than uniform weighting.

These measured operating points support the weighting choice in the evaluated setting. Since their achieved rates differ, we do not describe the row-wise differences as exact matched-rate gains or claim statistical significance. We will clarify the definition and MSE-only role of wiw_i in the loss description and add this ablation to the experimental discussion, while presenting feature norm as an empirically supported heuristic rather than a universal semantic-importance measure.

5. Figure 1, notation, and implementation description

We will revise Figure 1 and Section 3.1 together to ensure a one-to-one correspondence between the symbols in the figure and the text. We will use consistent module and variable names, define each signal at its first appearance, and add a caption legend for the attention and LoRA notation. The revised overview will read:

For each Gaussian, let xsemx_{\mathrm{sem}} denote its semantic feature and xattrx_{\mathrm{attr}} its appearance and geometric attributes. The projection encoder EprojE_{\mathrm{proj}} maps xsemx_{\mathrm{sem}} to the input space of the LoRA-adapted FCGS encoder Efea⊕E_{\mathrm{fea}}^{\oplus}, producing the semantic latent ysemy_{\mathrm{sem}}. GeoGuide GG uses the attribute-derived conditioning latent ygeoy_{\mathrm{geo}} to refine ysemy_{\mathrm{sem}} into y~sem\tilde{y}_{\mathrm{sem}}. The hyperprior encoder hah_a maps this refined latent to zz. Quantization QQ produces z^\hat{z} and y^\hat{y}, and the hyperprior decoder hsh_s uses z^\hat{z} to predict the probability-model parameters for y^\hat{y}. Finally, the LoRA-adapted decoder Dfea⊕D_{\mathrm{fea}}^{\oplus} and projection decoder DprojD_{\mathrm{proj}} reconstruct the semantic feature x^sem\hat{x}_{\mathrm{sem}} from y^\hat{y}. Here, ⊕\oplus denotes LoRA adaptation, and a hat denotes a quantized latent or reconstructed feature, as applicable.

6. Rendering FPS and computational overhead

In the separate implementation profile, after reconstructing and caching the per-Gaussian semantic features, the complete model renders RGB, 128-D features, and depth at 100.43 FPS, or 83.30 FPS when also including resize and the fixed 128→512 feature head. This is comparable to the uncompressed reference under the same settings.

Table R7. Separately profiled models on an NVIDIA A100-PCIE-40GB; these FPS measurements are not measurements of all five checkpoints in Table R1.

RepresentationRGB + 128-D features + depth (FPS)Including 512-D head (FPS)
Uncompressed reference100.7382.89
Complete model / LoRA100.4383.30
Same architecture / Full FT100.6683.21

Measurements use 990,975 Gaussians and 10 measured rounds per model and path, with synchronized wall-clock timing, batch size 1, and FP32. Raster outputs are 639×479; the 512-D feature output is 480×360. Timings exclude disk I/O, file export, text encoding, and semantic classification. They establish real-time rendering under this configuration, not end-to-end language-query throughput. The codec is run once per scene reconstruction rather than once per view; GeoGuide is encoder-side only.

Computational accounting. The shared 128→512 pointwise feature head costs 65.536 kMAC per semantic output pixel, excluding bias, for all three representations. Scene reconstruction is instead naturally measured per Gaussian. The following counts cover executed linear layers, not the full pipeline:

Table R8. Partial neural-operation counts; rasterization and non-linear operations are excluded.

Linear-layer computationLoRA MAC/GaussianFull FT MAC/Gaussian
Quantized reconstruction forward1,264,5121,199,104
Synthesis from cached quantized main latent393,216376,832

The corresponding measured whole-scene neural forward times are 0.438 s and 0.337 s; synthesis from cached main latents takes 0.090 s and 0.066 s. These measurements exclude actual entropy-bitstream decoding and disk I/O. We do not label them as full compression/decompression times, or combine the partial counts into a whole-pipeline KMAC/pixel estimate. The measured overhead supports separating one-time scene reconstruction from repeated rendering, and does not establish a computational speedup for unmerged LoRA.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than three days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in 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.