acceptodds
Withdrawn from ICLR 2027

PocketSplat: Mobile Gaussian Reconstruction via World-Space Latent Allocation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mobile Gaussian reconstruction must satisfy two requirements: the reconstruction model must execute within a device resource envelope, and the resulting Gaussian asset must expose a representation size suited to downstream mobile use. Existing feed-forward Gaussian reconstructors commonly decode dense, image-aligned candidates whose final cardinality is implicitly determined by the input resolution and number of views. We present PocketSplat, a feed-forward framework for budgeted mobile Gaussian asset construction. Given a prescribed output budget, PocketSplat organizes dense geometry-aware latent candidates in predicted world space, allocates exact integer capacity across local latent cells, and decodes complete Gaussian attributes only for retained candidates. Cell-conditioned latent fusion aggregates repeated multi-view evidence before decoding, while spatial responsibility decoding adapts Gaussian support after local sparsification. Experiments on DL3DV and out-of-distribution benchmarks establish a strong quality–budget trade-off against feed-forward Gaussian reconstruction baselines. On Mip-NeRF 360, PocketSplat executes directly on a target iPhone and constructs compact, higher-quality Gaussian assets substantially faster than a deployable streamed MVSplat variant; native MVSplat and DepthSplat exceed the device memory budget.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Feed-Forward 3D Reconstruction, Mobile 3D Vision, Budget-Aware Reconstruction, Geometry-Aware Allocation, Efficient Neural Rendering

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

This paper proposes PocketSplat, a budget-conditioned feed-forward framework for mobile 3D Gaussian reconstruction. The paper identifies that existing feed-forward Gaussian reconstructors typically produce dense, image-aligned candidates whose output cardinality is implicitly tied to input resolution and the number of views, making both device-side execution and downstream asset size difficult to control. PocketSplat instead formulates reconstruction as a world-space latent allocation problem with an explicit Gaussian-count budget. Dense geometry-aware candidates are grouped into predicted world-space cells, where an exact-budget allocator distributes integer capacity according to local visual detail and cross-view evidence. The retained candidates further benefit from cell-conditioned cross-view latent fusion, while Spatial Responsibility Decoding adjusts Gaussian scales according to local sparsification. Experiments show that PocketSplat achieves a quality–size trade-off and can run directly on an iPhone, demonstrating its potential for mobile deployment.

Critical strengths
  1. Exploration of mobile Gaussian reconstruction. The paper makes a meaningful exploration of feed-forward Gaussian reconstruction under mobile resource constraints. By introducing explicit budget control over the final Gaussian representation, it provides a practical direction for adapting Gaussian reconstruction to resource-limited on-device deployment.
Critical weaknesses
  1. Limited methodological novelty. The core design of organizing candidate elements into spatial cells and performing budget allocation at the cell level is conceptually similar to existing voxel-aligned or anchor-aligned representation paradigms. Although the paper combines this idea with cross-view evidence weighting, latent fusion, and scale adaptation for Gaussian reconstruction, the overall methodological novelty appears relatively limited. Several of these components can be viewed as adaptations or combinations of existing ideas, rather than a fundamentally new reconstruction paradigm.
  2. Insufficiently reasonable and comprehensive experimental evaluation. The experimental evaluation lacks comparisons with several recent state-of-the-art methods for reducing the number of Gaussians in feed-forward 3DGS, such as SparseSplat, Off-the-Grid, and EcoSplat. Moreover, the main experiments do not appear to follow the same experimental setup as the primary comparison method, F4Splat. This inconsistency in experimental settings makes the reported comparisons difficult to assess fairly and raises concerns about the reliability and reproducibility of the experimental conclusions. A more rigorous evaluation under consistent settings, together with comparisons against recent strong baselines, would be necessary to substantiate the claimed advantages.
Questions for authors
  1. Comparison with existing spatial representation paradigms. The proposed World-space Latent Allocation paradigm appears conceptually similar to voxel-aligned representations, such as AnySplat[1] and VolSplat[2], as well as anchor-aligned representations, such as AnchorSplat[3]. Could the authors clearly articulate the advantages of the proposed method over these existing paradigms, particularly in terms of budget controllability, reconstruction quality, and computational efficiency?
  2. Consistency of the experimental settings with F4Splat. The main baseline, F4Splat, is primarily evaluated on large-scale datasets such as RE10K and ACID under specific experimental settings. Could the authors evaluate PocketSplat on the same datasets and under the same experimental protocols to demonstrate that the observed performance improvements are robust and directly comparable?
  3. Unexpected performance of PocketSplat (Full). In Table 1, PocketSplat (Full) achieves a lower PSNR than PocketSplat (338K). This is somewhat counterintuitive, since increasing the number of Gaussian primitives would normally be expected to improve rendering quality. Could the authors provide a detailed explanation for this phenomenon and clarify what factors cause this behavior?
  4. Source of the low peak memory footprint. The paper does not clearly explain which specific design choices enable PocketSplat to maintain a low peak memory footprint. It would be helpful to provide a detailed analysis or ablation study showing the contribution of each component to memory efficiency, and to clarify whether this improvement mainly comes from the proposed representation, the budget-constrained decoding strategy, or other implementation details.
  5. Comparison with recent state-of-the-art methods. The evaluation would be more complete if the authors could compare the proposed method with recent state-of-the-art approaches targeting Gaussian count reduction or compact feed-forward 3DGS reconstruction, such as SparseSplat[4], Off-the-Grid[5], and EcoSplat[6].
  6. Generality and applicability of the proposed method. The current experiments are mainly conducted with 4 input views. It would be valuable to evaluate the generalization of PocketSplat to a larger number of input views, such as 8, 16, and 24 views, following the configurations of F4Splat. In addition, can the proposed framework be extended to pose-free settings, such as those explored by SPFSplat[7]?

[1] AnySplat: Feed-forward 3D Gaussian Splatting from Unconstrained Views. SIGGRAPH Asia 2025. [2] VolSplat: Rethinking Feed-Forward 3D Gaussian Splatting with Voxel-Aligned Prediction. ECCV 2026. [3] AnchorSplat: Feed-Forward 3D Gaussian Splatting with 3D Geometric Priors. CVPR 2026. [4] SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction. CVPR 2026. [5] Off The Grid: Detection of Primitives for Feed-Forward 3D Gaussian Splatting. CVPR 2026. [6] EcoSplat: Efficiency-controllable Feed-forward 3D Gaussian Splatting from Multi-view Images. CVPR 2026. [7] No Pose at All: Self-Supervised Pose-Free 3D Gaussian Splatting from Sparse Views. ICCV 2025.

Reviewer_hAKcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a feed‑forward framework for budget‑constrained mobile‑oriented 3D Gaussian reconstruction. Given a user‑specified primitive budget, the method lifts multi‑view observations into world‑space latent cells, performs exact‑capacity latent allocation guided by local visual details and cross‑view redundancy weighting, aggregates multi‑view evidence via hidden fusion, and adopts Spatial Responsibility Decoding (SRD) to adapt Gaussian covariance under spatial sparsification. Comprehensive experiments are conducted on DL3DV, Tanks & Temples, MegaDepth, together with real‑device deployment on iPhone using Mip‑NeRF 360. Empirical results show that PocketSplat achieves competitive quality‑budget trade‑offs and can run directly on mobile hardware, while existing feed‑forward baselines either suffer from out‑of‑memory failure or operate with much slower inference speed. Despite the promising technical pipeline and extensive benchmark evaluations, several critical concerns regarding motivation, novelty positioning and practical limitations prevent me from recommending acceptance.

Critical strengths
  1. Thorough evaluation protocol: The experimental evaluation is comprehensive. The authors conduct in‑domain and out‑of‑distribution quantitative benchmarks, qualitative visual comparisons, controlled ablations to validate core components (world‑space allocation, SRD, allocate‑before‑decode order), GPU runtime profiling, and actual end‑to‑end iPhone deployment. Ablations clearly demonstrate performance gains brought by each proposed module.
  2. This work starts from a well‑motivated goal: enabling mobile‑friendly 3D Gaussian reconstruction by explicitly controlling the number of Gaussian primitives, and it delivers a complete end‑to‑end pipeline toward this target.
  3. Successful On-Device Deployment: The proposed framework can run directly on an iPhone, demonstrating its practical feasibility for mobile 3D Gaussian reconstruction. It achieves end-to-end reconstruction in approximately 10 seconds, highlighting its potential for real-world mobile applications.
Critical weaknesses
  1. Regarding experimental setup: standard NVS benchmarks such as DL3DV typically use camera trajectories orbiting a central object. The authors do not specify how input views are selected in their experiments. From the provided figures, view sampling appears somewhat random. Under such random view selection, depth-dependent methods such as DepthSplat may fail to produce reliable depth estimates, leading to severe performance degradation. This raises two concerns. First, it is unclear whether the selected input views are reasonable and practically meaningful. Proper reconstruction should inherently rely on well-chosen views. Both DepthSplat and MVSplat uniformly sample frames along the capture trajectory, which represents a reasonable and practically relevant setup. Second, poor selection of the four input views can lead to degraded performance across all methods, rendering the experimental results less meaningful. For instance, in the truck scene from Tanks & Temples in Figure 4, all methods perform poorly. Even though the proposed method achieves slightly better results, the comparison lacks convincing power. The authors should adopt appropriate view sampling protocols.

  2. ResSplat can achieve reconstruction with fewer Gaussians and support higher-resolution rendering, so the authors should include it in comparisons. In addition, token-based methods such as TokenGS and its extension TokenGS+Anchor (LocusGS: Spatially Grounded Tokens for Feed-Forward 3D Gaussian Splatting) belong to the same category as the proposed approach and should also be evaluated.

  3. The visual results provided by the authors are not sufficiently promising. The flawed experimental setting and non-standard view selection undermine the quality of visualizations.

  4. Insufficient isolation of the proposed method's contributions from the geometry backbone. The proposed method relies on a frozen pretrained multi-view geometry backbone to provide depth estimates, camera parameters, and intermediate features, while the compared methods employ different architectures and geometric priors. Therefore, the substantial performance gains reported in Table 1 cannot be confidently attributed to the proposed world-space allocation and Gaussian decoding strategies. The authors should provide controlled comparisons using the same geometry backbone and training protocol, or introduce a baseline that directly predicts Gaussians from the same backbone without the proposed allocation mechanism. Such experiments are necessary to establish whether the improvements primarily arise from the proposed method rather than stronger pretrained geometric features.

  5. Limited computational efficiency gains from budget-controlled reconstruction. Although the proposed framework explicitly controls the number of output Gaussians and emphasizes allocating candidates before complete attribute decoding, the reported efficiency results raise concerns about its practical computational benefits. According to Table 6, reducing the Gaussian budget from 451K to 113K actually increases GPU inference time from 322.3 ms to 601.8 ms.

Questions for authors

please refer to the weaknesses

Reviewer_c6gcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PocketSplat addresses feed-forward Gaussian reconstruction with a user-specified output budget and direct mobile execution. A frozen multi-view geometry backbone produces dense geometry-aware latent candidates, which are grouped into world-space cells. The method assigns exact integer cell quotas, fuses information within cells, and decodes retained candidates into Gaussians. Spatial responsibility decoding (SRD) adjusts Gaussian scales according to local compression, with learned corrections. Evaluation covers reconstruction quality on DL3DV and out-of-distribution datasets, component ablations, GPU profiling, and an iPhone deployment compared with a streamed MVSplat variant.

Critical strengths
  • The problem is practically relevant, and the system offers explicit output-size control. Producing multiple prescribed Gaussian budgets with one deployed model is useful for mobile applications. The quality–budget results and direct iPhone execution provide meaningful evidence of practical feasibility.

  • The SRD ablation provides concrete evidence for support adaptation. Table 3 reports gains of approximately 3.36 dB at 113K and 0.83 dB at 226K, using branches trained from a common parent checkpoint and positive paired bootstrap confidence intervals. This supports the value of compensating for support lost during sparsification, particularly at low budgets.

Critical weaknesses
  • The distinct technical advance over existing budget-controlled reconstruction methods is not sufficiently established. F⁴Splat already addresses budget-controlled feed-forward Gaussian prediction, and world-space grouping and feature aggregation have related precedents. PocketSplat differs in its exact cell quotas and their combination with latent fusion and compression-conditioned scale adaptation. However, the paper does not clearly establish what new capability or insight this combination provides beyond existing structured allocation and aggregation strategies. Exact cardinality control and mobile execution are useful, but the added value of the proposed algorithmic combination needs stronger evidence. My concern is the insufficiently established contribution, rather than the use of familiar components itself.

  • The experiments do not adequately isolate the proposed mechanisms or explain the sources of the gains. Table 3 compares the allocator mainly against uniform random selection, which is insufficient to establish the advantage of the particular world-space quota policy over other structured allocation policies. Cell-conditioned latent fusion is not separately ablated, nor are the contributions of the entropy-based quotas and view-mean normalization isolated. The SRD ablation is meaningful, particularly at low budgets, but it compares learned adaptation with no adaptation; a comparison against the fixed rule already contained in its formulation, �0��0�, is needed to assess the added value of learning. At Full, where the allocator retains all candidates, PocketSplat still exceeds F⁴Splat by approximately 4.47 dB on DL3DV (Table 1). Thus, the inter-method comparison alone cannot identify how much improvement comes from allocation rather than the backbone, Gaussian head, fusion, training, or scale correction. More clearly specified baseline configurations and targeted controls are needed. The definition of �1��1� and the training roles of reliability/support predictions are also unclear. Together, these omissions leave the method's central design choices insufficiently validated.

  • Minor reporting comment: Please reconcile the text–table discrepancies: Section 4.2 states that 21.018 dB exceeds pixelSplat's 21.231 dB; Section 4.3 calls Quality's SSIM 0.6503 the best, although Light has 0.6690 in Table 2. Table 6's 17.796 GB peak also needs reconciliation with the stated 16 GB RTX 4080 SUPER. These are supplementary clarification points, not the main basis of my recommendation.

Questions for authors
  • Contribution and allocation: What is the principal technical advantage of the exact world-space quota policy over existing structured budget-controlled prediction? Can you provide a structured allocation control using the same backbone/head, and isolate the effects of the entropy-based quotas and view-mean normalization?

  • Fusion and SRD: Can you separately ablate cell-conditioned fusion and compare learned SRD with the fixed �0��0� scale rule, holding the budget and training protocol fixed? These targeted controls would establish whether the reported improvements require the proposed mechanisms.

  • Baseline attribution and training: Please specify the checkpoints, training data, camera inputs, and budget settings used for F⁴Splat and PocketSplat. How should the approximately 4.47 dB gap at Full, where no candidates are removed, be interpreted? Please also define �1��1� and explain the supervision/gradient paths for reliability, support, and allocation-related modules.

  1. Brief reporting clarification: Please correct or explain the text–table discrepancies noted above and clarify the hardware or memory statistic underlying Table 6's 17.796 GB entry.

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

Related papers

Discussion (0)

Sign in to comment.