acceptodds
Under review as a conference paper at ICLR 2027

ZoomSplat: Rethinking Multi-view Frequency Bounds for Anti-aliased Feed-forward 3D Gaussian Splatting

Abstract

Feed-forward 3D Gaussian Splatting (FF-3DGS) enables efficient sparse-view reconstruction and novel view synthesis, but these reconstructions can exhibit severe erosion and grid-like artifacts when rendered beyond the context-image sampling rates such as zoom-in operation. We propose ZoomSplat for FF-3DGS with detail-preserving reconstruction and smooth higher-resolution rendering from a fixed Gaussian representation. Although FF-3DGS models combine information from multiple input views, existing multi-view 3D frequency constraint governed by the largest sampling rate of individual views does not account for their combined observations and can impose excessive smoothing. To address this, our ZoomSplat integrates multiview-aware 3D filtering into Gaussian prediction for the first time, through an aggregated multiview frequency bound that permits finer Gaussian scales in regions supported by multiple observations. Explicit filter-size conditioning then enables the FF-3DGS models to account explicitly for the resulting filter sizes. Finally, an Opaque Scene Prior (OSP) combines termination-conditioned color normalization with transmittance regularization to compensate for residual transparency and to encourage opaque coverage. On DL3DV with six context views, ZoomSplat improves average 1.4dB PSNR across , , , and rendering scales over TokenGS that is the second best FF-3DGS model and lifts average 3.2dB PSNR over the baseline of DepthSplat.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Novel View Synthesis, Feed-forward 3D Reconstruction, Anti-aliasing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_79QSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GeoGSplat, a geometry-grounded feed-forward 3D Gaussian Splatting (3DGS) framework that operates in a non-pixel-aligned paradigm. While existing pixel-aligned methods suffer from a primitive count that scales linearly with input resolution and view count (N=NI×H×WN = N_I \times H \times W), recent non-pixel-aligned methods predict a fixed number of Gaussians via learnable tokens but discard explicit ray/depth geometric constraints, leading to poorly placed Gaussian centers and severe rendering artifacts under large viewpoint changes.

Critical strengths

Well-Motivated Hybrid Paradigm Resolving the Efficiency-Geometry Trade-off: The core diagnosis of why non-pixel-aligned 3DGS methods (like TokenGS) fail under large viewpoint changes—namely, photometric overfitting due to underconstrained Gaussian centers—is insightful and empirically well-substantiated (especially in Section 4.3, Figures 5 & 6, and Appendix A.1). By coupling fixed-budget Anchor queries with continuous Triplane feature fields, GeoGSplat maintains a constant primitive count (65K65\text{K}) regardless of whether 2, 4, or 12 views are input, while achieving superior geometric accuracy (Chamfer Distance reduced from 0.2770.277 to 0.0170.017 on NRGBD compared to TokenGS) and significantly better extrapolation to distant target views. A common pitfall in large-scale feed-forward 3D vision papers is confounding architectural gains with dataset scaling or supervision differences.

Critical weaknesses

Joint self/cross-attention over M=4096M=4096 Anchor queries and nT=3×642=12,288n_T = 3 \times 64^2 = 12,288 Triplane queries across 12 shared attention blocks.

Two stages of Triplane upsampling up to 3×256×2563 \times 256 \times 256 with 3D-to-2D cross-plane fusion and image-conditioned cross-attention at high spatial resolutions (Eq. 18).

Computing symmetric Chamfer Distance and Farthest Point Sampling (FPS) on 65,53665,536 points during training (requiring 128 NVIDIA L20 GPUs). Without wall-clock runtime and memory metrics, the claim of "computational efficiency" remains only half-proven (rendering is fast, but encoding/decoding might be a bottleneck).

Questions for authors

Could the authors provide the core specifications deferred to Appendices B, F, H, and I? Specifically: (a) What is the exact base VLM backbone (e.g., Qwen3-VL size)? (b) What are the dataset statistics (number of wafers, lots, operations, classes, and CF-RC test pairs)? (c) What are the details of Table 7 (the 4-phase training schedule) and the computational overhead of multi-attempt rollouts (K≤3K \le 3) plus the clue-conditioned forward pass relative to standard GRPO?

Equations (4) and (7) bound the Anchor centers and Triplane queries to [−1,1]3[-1, 1]^3 via tanh⁡(⋅)\tanh(\cdot). In Appendix B.5, you mention that during geometric evaluation, "the ground-truth point cloud determines its scale" to normalize the scene into [−1,1]3[-1, 1]^3. During inference (before GT geometry is known), how is the scene coordinate frame and scale defined solely from the input camera poses πi\pi_i so that unbounded scenes (like Mip-NeRF 360 in Appendix A.3) fit properly inside the [−1,1]3[-1, 1]^3 Triplane bounds without clipping background geometry?

Why did you choose to compute Lrender\mathcal{L}_{render} only on the 4 input views during training rather than held-out target views (which is standard in feed-forward NVS)? Does supervising on novel target views further improve PSNR/LPIPS, or allow the model to rely less on GT depth supervision (Lgeo\mathcal{L}_{geo})?

In Eq. (8), ΔXi=CrossAtt(Qioff,Figeo,Figeo)\Delta X_i = \text{CrossAtt}(Q_i^{off}, F_i^{geo}, F_i^{geo}). Are the learnable offset tokens Qoff∈RK×CQ^{off} \in \mathbb{R}^{K \times C} shared across all MM Anchors, or distinct per Anchor? Moreover, since Figeo∈RK×3CTF_i^{geo} \in \mathbb{R}^{K \times 3C_T} already has a 1-to-1 correspondence with the KK sampled local reference points Xˉi\bar{X}_i, why use an all-to-all K×KK \times K cross-attention with learnable queries instead of a point-wise MLP or local self-attention directly on FigeoF_i^{geo}?

Reviewer_3uDHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ZoomSplat addresses artifacts that arise when feed-forward Gaussian reconstructions are rendered at resolutions above those of the input images. Built on DepthSplat, it replaces the maximum per-view sampling-rate rule with an aggregation of squared sampling rates to determine Gaussian smoothing scales. It additionally conditions Gaussian attribute prediction on the filter size and introduces an opaque-scene prior combining accumulated-alpha color normalization with transmittance regularization. The model is trained with low-resolution supervision and reuses a fixed Gaussian representation across output resolutions. Evaluation on 139 DL3DV test scenes and two Tanks and Temples scenes reports improved image quality across rendering scales, with component ablations and additional input-view settings.

Critical strengths
  1. The paper addresses a concrete limitation of feed-forward reconstruction. Artifacts exposed by increased rendering resolution affect the usability of reconstructed scenes. Improving image quality while retaining a fixed Gaussian representation and using only low-resolution training supervision is practically valuable.
  2. The empirical evaluation demonstrates benefits beyond the original backbone. The paper evaluates 139 held-out DL3DV scenes, several input-view counts, and zero-shot transfer to two Tanks and Temples scenes. Comparisons include recent feed-forward methods and simple render-then-upsample alternatives. The consistent improvements across image-quality metrics support the usefulness of the complete system.
Critical weaknesses
  1. The central multiview frequency aggregation is insufficiently justified relative to simpler filtering alternatives. The proposed aggregation increases the effective sampling rate even for redundant observations, without accounting for sample arrangement, surface orientation, or occlusion. The paper acknowledges these limitations, but does not establish when the rule captures complementary information rather than simply reducing smoothing strength. In Table 3, replacing conventional filtering with MAS3DF improves average PSNR by approximately 0.14 dB, whereas introducing conventional filtering already yields a 2.40 dB gain. A comparison against a tuned maximum-rate filter, with OSP and conditioning held fixed, is needed to establish the significance of the central contribution.
  2. The evaluation does not cleanly isolate the claimed high-resolution detail benefits. ZoomSplat already improves native-resolution reconstruction, while OSP can improve pixel fidelity by correcting opacity-related darkening. Comparisons against upsampled DepthSplat renderings therefore mix baseline reconstruction quality with benefits specific to higher-resolution rendering. The missing comparison against upsampling ZoomSplat’s own native-resolution outputs leaves uncertainty about how much additional detail direct high-resolution rendering preserves. Furthermore, increasing output resolution does not establish robustness to camera approach, which also changes visibility and perspective. These distinctions limit the broader interpretation of the reported gains.
Questions for authors
  1. Targeted validation of the aggregation rule: With OSP, conditioning, and the training protocol held fixed, how does MAS3DF compare with the maximum-rate rule using a tuned filtering strength? Does the advantage persist when context views are highly redundant, rather than providing complementary observations?
  2. Targeted validation of high-resolution rendering: How does directly rendering ZoomSplat at 2×, 4×, and 8× compare with bicubic upsampling its own 1× renderings? This would separate improvements in the underlying reconstruction from benefits specific to rendering at increased sampling rates.
  3. Clarification of evaluation-time opacity handling: How were the accumulated-alpha thresholds of 0.1 during training and 0.001 during evaluation selected? How sensitive are the reported results to this change, and were these choices determined using a validation set separate from the reported test scenes?
Reviewer_YC9wOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ZoomSplat, a Anti-aliased Feed-forward 3D Gaussian Splatting method with powerful performance. The paper is well-organized. This paper focuses on the problem that previous methods does not account for their combined observations and they are easy to produce some dark artifacts. Therefore, this paper proposes multiview-adaptive sampling-rate-based 3D filtering with an opaque-scene prior and filter-size conditioning to suppress artifacts. The experiments demonstrate the performance of the proposed method.

Critical strengths
  1. The language is well-structured and this paper is easy to follow.

  2. The proposed method performs well in various sampling rates across differernt datasets.

Critical weaknesses
  1. The authors propose multiview-adaptive sampling-rate-based 3D filtering (MAS3DF) to address the problem that S3DFonly selects a single per-view rate. In my opinon, the authors seem only change the original single-view approach to multi-view and multiple rates. MV methods can definitely beat the single-view paradigm. Therefore, I did not get any insight from this idea.

  2. The proposed Opaque-scene prior seems promising and it is the most important part of the proposed method to improve performance. I am curious about this strategy. The problem is how the authors get the prior? Do they obtain the idea from SurfSplat? I am not sure whether this strategy could change the distribution of opacity. It would be bettter to show that. Also, this strategy may push the opaicty toward 1. In this case, there may some solid floaters.

  3. Is the propose method directly producing the Gaussians map without any other fine-tune for Gaussians? What about the inference time compared with other methods?

I might improve my score after rebuttal.

Questions for authors
  1. Any plan to release the code and weights?

  2. May I see the distribution of the opacity and scale from the produced Gaussians? It would be better the show hist; if the rebuttal cannot provide it, please type it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.