acceptodds
Under review as a conference paper at ICLR 2027

CSplat: Exploiting Cross-View Redundancy for Compact Feed-Forward 3D Gaussian Splatting

Abstract

Recent feed-forward 3D Gaussian Splatting (3DGS) compression methods achieve substantial rate-distortion gains over directly compressing large sets of unordered Gaussian primitives by coding multi-view intermediate features. However, these view-wise latent streams are still entropy-coded independently, overlooking cross-view redundancy and its strong dependence on inter-view geometry. We introduce CSplat, a Cross-view Causal Compression framework for Gaussian Splatting, which formulates feature compression in feed-forward 3DGS as cross-view causal entropy modeling. Each view is encoded using only previously decoded views and decoder-available geometry. CSplat aligns decoded reference features to the current view using predicted depth and camera geometry, adaptively aggregates references according to spatial reliability, and combines the resulting cross-view condition with the existing single-view hierarchical prior to refine the probability distribution used for arithmetic coding. On the 8-view DL3DV-Benchmark, CSplat achieves a PSNR BD-rate of over the state-of-the-art method. We further find that bitrate reduction is strongly correlated with geometric view overlap, confirming that cross-view redundancy is closely tied to shared visible scene content and remains exploitable even after multi-view interaction in the reconstruction backbone. These results establish cross-view causal entropy modeling as an effective mechanism for exploiting cross-view redundancy toward more compact feed-forward 3DGS representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Feed-Forward 3D Gaussian Splatting, 3D Gaussian Splatting Compression, Cross-View Entropy Modeling, Learned Compression, Multi-View Geometry

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uY8POpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes C$^3$Splat, a framework that exploits cross-view redundancy in latent feature compression for feed-forward 3DGS. The method is built upon CodecSplat, and further introduces causal cross-view context modeling to improve entropy coding efficiency. Specifically, it uses predicted depths and camera parameters at the decoder side to warp previously decoded features into the target view, followed by reliability-aware aggregation to construct cross-view contexts. These contexts are then used to refine the hierarchical entropy prior for more accurate probability estimation. Experiments on the 2-view RE10K and 8-view DL3DV benchmarks demonstrate improved rate-distortion performance over baseline methods. The paper further analyzes the relationship between geometric overlap and bitrate reduction, showing that higher overlap generally leads to greater coding gains.

Critical strengths
  1. Efficient design and easy integration: The proposed method leverages predicted depth and camera geometry to construct cross-view contexts through simple feature warping, offering a more straightforward and geometrically grounded alternative to other complex learned cross-view interaction mechanisms. Moreover, the proposed cross-view branches can be readily integrated into pretrained compression models with minimal architectural modifications.

  2. Effective experimental validation: The paper provides extensive experiments on both 2-view and 8-view benchmarks, demonstrating the effectiveness of cross-view context modeling under different settings. In particular, the entropy-only adaptation experiments isolate the bitrate savings from improved probability estimation without affecting reconstruction quality. The additional analysis of geometric overlap further supports the motivation and effectiveness of the proposed method.

Critical weaknesses
  1. Limited evaluation of decoding complexity: The proposed cross-view context modeling introduces additional geometric warping, multi-reference aggregation, and entropy refinement operations. However, the paper only provides theoretical complexity analysis without reporting actual runtime measurements, making it difficult to assess the practical decoding overhead of the proposed method.

  2. The ablation studies could be deeper: Although the paper provides ablation studies on several key components, the experimental analysis remains somewhat limited. For example, the effectiveness of the proposed causal view ordering strategy is not sufficiently investigated. The impact of the number of input views on compression performance also remains unclear.

Questions for authors
  1. Could the authors compare different view ordering strategies, such as random ordering and overlap-aware ordering, to justify the proposed strategy or evaluate the robustness of the method to different decoding orders?

  2. Could the authors provide additional experiments with different numbers of input views on the same dataset to investigate how the compression gains scale with the number of available reference views? If supported by the pretrained models, an evaluation using the original CodecSplat and entropy-only C³Splat checkpoints without additional training could provide a lightweight analysis of this behavior.

2 comments

1. Practical Encoding and Decoding Complexity

We agree that actual runtime measurements are important for assessing the practical overhead of cross-view causal compression. We have therefore added encoding and independent decoding benchmarks across 2, 4, 8, and 16 input views, reporting runtime and peak allocated GPU memory in Appendix F (Table 10). The experiments use a deployed compression pipeline in which the encoder generates a real scene bitstream, and a separate decoder reconstructs Gaussians using only the bitstream and model parameters. On a single RTX 4090 with eight input views at λ = 256, C³Splat requires 1.018 s for encoding and 0.806 s for decoding, compared with 0.507 s and 0.361 s for CodecSplat, while introducing approximately 0.08 GiB of additional peak GPU memory at each endpoint. As analyzed in Section 4.6, the computational overhead mainly arises from constructing cross-view conditions for previously decoded reference–target pairs, whose number grows quadratically with the input-view count. However, aggregating references before condition encoding avoids repeated condition-encoder computation, and the implementation optimizations described in Appendix E further reduce unnecessary overhead without changing the probability model or bitstream. These measurements clarify the trade-off between improved compression efficiency and increased decoding latency, which we have also explicitly acknowledged in the revised manuscript.

2. View-ordering Strategies

Following the reviewer's suggestion, we have added a dedicated experiment in Section 5.3 (Table 3) comparing five decoding strategies: temporal forward, temporal reverse, medoid + random, fully random, and our proposed medoid + nearest-set ordering. All five strategies achieve the same reported PSNR at each operating point, while our ordering consistently yields the smallest bitstreams and the best overall PSNR BD-rate of −20.96% relative to CodecSplat. Temporal orderings remain close, with bitrate increases of only 0.07%–0.34%, whereas random orderings result in increases of 0.67%–1.31%. These relatively modest differences indicate that the compression gains are robust to different decoding orders, rather than relying on a particular view sequence. The ability to aggregate multiple decoded references may help reduce sensitivity to individual reference choices, while geometry-based ordering provides a consistent additional benefit. We have also clarified the deterministic view-selection procedure in Section 4.4 and Appendix E.1, which allows the encoder and decoder to reconstruct the same ordering without transmitting additional information.

Table 3. View-ordering comparison on DL3DV-Benchmark.

View orderingλ = 64 Rate (kB) ↓λ = 256 Rate (kB) ↓λ = 1024 Rate (kB) ↓Overall BD-rate ↓
Temporal forward34.7762.3690.17−20.91%
Temporal reverse34.8662.3589.95−20.71%
Medoid + random34.9862.7690.64−20.44%
Fully random35.0963.0191.04−20.30%
Medoid + nearest-set (ours)34.7562.2489.86−20.96%

3. Effect of Input-view Count

We appreciate the suggestion to investigate how compression gains scale with the number of available reference views. We have added experiments with 2, 4, 8, and 16 input views on DL3DV-Benchmark, as reported in Section 5.4 (Figure 5) and Appendix H. For each view count, we independently trained a CodecSplat baseline and initialized the corresponding C³Splat model from that checkpoint, performing entropy-only adaptation while keeping the quantized latents and reconstruction path fixed. This ensures that the observed bitrate reductions arise from improved probability modeling rather than changes in reconstruction quality within each view-count setting. As shown below, at λ = 1024, the bitrate savings increase from 4.55% with 2 views to 23.67% with 16 views. Across all seven rate-control operating points, the corresponding PSNR BD-rate reductions are −1.77%, −5.72%, −12.00%, and −16.92%, respectively. These results show that C³Splat benefits increasingly from additional reference views, consistent with the availability of more shared scene information for cross-view entropy conditioning. Together with the geometric-overlap analysis in Section 5.4 (Figure 4), these findings provide further evidence that the compression improvements are associated with exploitable cross-view redundancy.

Input-view Count Analysis on DL3DV-Benchmark (λ = 1024).

Input viewsCodecSplat Rate (kB) ↓C³Splat Rate (kB) ↓Rate Saving ↑PSNR BD-rate ↓
237.3635.664.55%−1.77%
465.2158.1410.84%−5.72%
8105.8987.9916.90%−12.00%
16180.21137.5623.67%−16.92%

Thank you for your response. The additional experiments further demonstrate the robustness of the proposed method and have addressed my concerns. I will maintain my positive rating.

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

This paper addresses feed-forward 3DGS compression, where the existing method CodecSplat entropy-codes each view's latent independently and ignores redundancy across views. C³Splat codes views one by one in a deterministic causal order. It uses the features, predicted depths, and camera parameters of already decoded views, aligns them to the current view via forward splatting, aggregates them with per-pixel reliability weights, and uses the result to refine the mean and scale of the single-view hierarchical prior.

Critical strengths

The problem is clearly identified and well motivated, since the per-view latents in feed-forward 3DGS naturally contain cross-view redundancy. The method relies only on information available at the decoder, so it adds no side information and does not need to transmit a reference graph.

Critical weaknesses
  1. The evaluation is small in scale and lacks practical analysis. Each dataset has only 29 test scenes, and RE10K uses only 2 views.
  2. The comparison is not comprehensive. The only meaningful comparison is with CodecSplat; there is no comparison within the same framework against cross-view entropy models from multi-view image compression (e.g., epipolar attention or VEMIC-style methods). This makes it hard to show that geometric alignment is better than generic attention.
Questions for authors

Please refer to the weakness part.

2 comments

1. Expanded Evaluation Scale and Input-view Count Analysis

We agree that a larger evaluation scale is important for assessing the effectiveness of our method. In the revised manuscript, we have expanded the evaluation from the original 29-scene subset to the full 140-scene DL3DV-Benchmark, following the ReSplat evaluation protocol, and updated all corresponding experimental results (Section 5.1). To further investigate how compression performance scales with the number of available reference views, we conducted additional experiments with 2, 4, 8, and 16 input views on DL3DV-Benchmark (Section 5.4, Figure 5, and Appendix H). For each view count, we independently trained the corresponding CodecSplat baseline and performed entropy-only adaptation, keeping the quantized latents and reconstruction outputs unchanged. The resulting PSNR BD-rate reductions increase from −1.77% with 2 views to −5.72%, −12.00%, and −16.92% with 4, 8, and 16 views, respectively. These results provide further evidence that C³Splat effectively exploits cross-view redundancy and achieves increasing compression gains as more reference views become available.

Input-view Count Analysis on DL3DV-Benchmark (λ = 1024).

Input ViewsCodecSplat Rate (kB) ↓C³Splat Rate (kB) ↓Rate Saving ↑PSNR BD-rate ↓
237.3635.664.55%−1.77%
465.2158.1410.84%−5.72%
8105.8987.9916.90%−12.00%
16180.21137.5623.67%−16.92%

2. Comparisons with Cross-view Compression Methods

We agree that comparisons with existing cross-view entropy modeling mechanisms are important for evaluating the effectiveness of our design. Following the reviewer's suggestion, we introduced two additional baselines within the same CodecSplat framework: CodecSplat + Epipolar Attention, adapted from ECSIC, and CodecSplat + VEMIC, which incorporates view-aware conditioning over previously decoded latents. Both baselines preserve the original FF3DGS reconstruction pipeline and causal decoding constraints. Their detailed architectures and implementation procedures are provided in Appendix D, while parameter counts and runtime comparisons are reported in Appendix E. As shown in the revised Table 1, both approaches improve upon CodecSplat, confirming the benefits of incorporating cross-view information into entropy modeling. However, C³Splat achieves substantially greater compression efficiency: at λ = 1024 on DL3DV-Benchmark, our method achieves 89.86 kB / 27.34 dB, compared with 102.73 kB / 27.24 dB for Epipolar Attention and 103.49 kB / 27.21 dB for VEMIC. These comparisons provide direct evidence that our geometry-aligned causal conditioning is more effective than the evaluated adaptations of existing multi-view interaction mechanisms within the same FF3DGS compression framework.

Cross-view Compression Comparisons on DL3DV-Benchmark (λ = 1024).

MethodRate (kB) ↓PSNR (dB) ↑SSIM ↑LPIPS ↓
CodecSplat105.8927.150.8910.105
CodecSplat + Epipolar Attention102.7327.240.8920.101
CodecSplat + VEMIC103.4927.210.8920.102
C³Splat (ours)89.8627.340.8920.101

3. Practical Encoding and Decoding Analysis

We agree that practical efficiency is an important consideration for compression methods. We have therefore added actual encoding and independent decoding benchmarks across 2, 4, 8, and 16 input views, reporting runtime and peak allocated GPU memory in Appendix F (Table 10). The benchmark uses a complete deployed compression pipeline: encoding maps posed context images to a real scene bitstream, while a separate decoding process reconstructs Gaussians using only the bitstream and model parameters, without retaining input images or encoder-side intermediate features. Both methods are evaluated under the same hardware and timing protocol, with identical input-view selections. On a single RTX 4090 with eight input views at λ = 256, C³Splat requires 1.018 s for encoding and 0.806 s for decoding, compared with 0.507 s and 0.361 s for CodecSplat, while introducing only approximately 0.08 GiB of additional peak GPU memory at each endpoint. As explained in Section 4.6 and Appendix F, the additional computation arises from constructing causal cross-view conditions, with the number of reference–target pairs growing quadratically with the input-view count. Aggregating references before condition encoding reduces repeated condition-encoder computation, while the implementation optimizations detailed in Appendix E avoid unnecessary overhead without changing the transmitted bitstream or probability model. These measurements clarify the practical trade-off: C³Splat achieves improved compression efficiency at the cost of increased encoding and decoding time, while maintaining memory consumption close to CodecSplat.

Reviewer_MEaSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses compression in the context of feed-forward 3DGS.Current methods neglect cross-view redundancy and its strong dependence on inter-view geometry. To address this issue, the authors propose a Cross-view Causal Compression framework. Its core concept is to align decoded reference features with the current view using predicted depth and camera geometry information, adaptively aggregate reference information based on spatial reliability, and combine the resulting cross-view conditions with existing single-view hierarchical priors to improve the probability distribution used for arithmetic coding.

Critical strengths

Cross-view Causal Compression framework is used to boost the coding efficiency of CodecSplat which ignore the inter-view redundancy using entropy coding.

Critical weaknesses
  1. This work is built upon on CodecSplat and the main modification is to improve the entropy model by considering the inter-view redundancy. Novelty is limited. The idea of using inter-view features in the context model of the current view is straightforward. 2.CodecSplat is not introduced with sufficient detail, nor is the proposed method. Authors omit many indices for simplicity concern, but it may lead to clarity issues.
  2. How is the spatial context integrated with the inter-view context?
  3. Missing of the baseline for comparison, particularly those FF3DGS compression techniques considering multi-view feature compression.
Questions for authors

(2) needs more explanation. Figure 1 is introduced without sufficient description in the paragraph. Ablation study for view order determination.

1 comment

2. Technical Presentation and Context Integration

We have substantially revised Sections 3 and 4 to improve the technical clarity of the manuscript. Specifically, Appendix A now provides a detailed introduction to CodecSplat, including its multi-view feature extraction, pre- and post-codec refinement, hierarchical feature compression, and Gaussian reconstruction pipeline, making the distinction between the original framework and our contributions clearer. We have also expanded the description of Figure 1 in Section 4.1 and clarified the mathematical notation and indices throughout the method section. Regarding the integration of spatial and cross-view contexts, Section 4.3 now explicitly explains how the cross-view conditions refine the mean and scale parameters predicted by CodecSplat's hierarchical prior through residual corrections, rather than replacing the original prior. Appendix B further details the coder-aligned mean–scale parameterization and arithmetic coding procedure, while Appendix C provides the network architectures and implementation details. Additionally, Appendix E describes the complete causal encoding and decoding procedures, including how encoder–decoder consistency is maintained. We believe these revisions provide a clearer and more comprehensive explanation of the proposed method.

3. View-ordering Ablation

Following the reviewer's suggestion, we have added a dedicated view-ordering experiment in Section 5.3 (Table 3), comparing five strategies: temporal forward, temporal reverse, medoid + random, fully random, and our proposed medoid + nearest-set ordering. All five strategies achieve the same reported PSNR at each operating point, while our strategy consistently produces the smallest bitstreams and achieves the best overall PSNR BD-rate of −20.96% relative to CodecSplat. Temporal orderings remain close, with rate increases of only 0.07%–0.34%, whereas random orderings increase the rate by 0.67%–1.31%. These results suggest that the compression gains are robust to view order, while geometry-based ordering provides a consistent additional benefit. We have also clarified the deterministic view-selection procedure in Section 4.4 and its implementation in Appendix E.1, explaining how the encoder and decoder reconstruct the same ordering from camera parameters without transmitting additional ordering information.

Table 3. View-ordering comparison on DL3DV-Benchmark.

View orderingλ = 64 Rate (kB) ↓λ = 256 Rate (kB) ↓λ = 1024 Rate (kB) ↓Overall BD-rate ↓
Temporal forward34.7762.3690.17−20.91%
Temporal reverse34.8662.3589.95−20.71%
Medoid + random34.9862.7690.64−20.44%
Fully random35.0963.0191.04−20.30%
Medoid + nearest-set (ours)34.7562.2489.86−20.96%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.