acceptodds
Under review as a conference paper at ICLR 2027

TileSkipper: Region-Adaptive Tile Pruning for 3D Gaussian Splatting

Abstract

Tiled 3D Gaussian Splatting rasterizers often use one scene-wide contribution cutoff for tile enumeration, although content differs in its sensitivity to support truncation. TileSkipper selects a static per-Gaussian cutoff policy for a frozen checkpoint. Calibration renders measure candidate pair savings and an isolated-removal distortion proxy that accounts for front transmittance and background color. The method allocates cutoffs across 64 Gaussian groups and accepts policies only after complete renders on disjoint selection views. The exported policy uses one byte per Gaussian, with no parameter updates, additional kernel, or per-frame policy inference. Across 13 scenes from Mip-NeRF 360, Tanks & Temples, and Deep Blending, a fixed-policy AccuTile sweep gives dataset-macro speedups of at standard resolution and at 3840 pixels wide, with dB mean PSNR change. Six integrations with existing opacity-aware bounds yield – compiler-only speedups. For four ports from rasterizers, we separately attribute the prior exact-bound transition and our incremental gain. Matched-quality ablations show modest gains over scene-global calibration and parity with per-Gaussian control; the standalone comparison with AdaGScale is regime-dependent.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Neural Rendering, Novel View Synthesis, Rendering Acceleration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RRVHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TileSkipper, a region-adaptive tile pruning method for accelerating 3D Gaussian Splatting. It uses offline calibration to assign different contribution thresholds to Gaussian groups, reducing Gaussian-tile pairs without modifying the original Gaussians. Experiments across 13 scenes and multiple rasterizers show modest rendering speedups with little impact on rendering quality.

Critical strengths

The main strength of this paper is its thorough and transparent experimental evaluation. The authors provide extensive ablation studies, comparisons across different rasterizers, and detailed performance breakdowns. They also clearly separate the speedup from existing techniques and the actual contribution of their method, and openly discuss its limitations and negative results. This makes the experimental results more convincing and easier to assess.

Critical weaknesses
  1. Limited Practical Speedup The actual rendering speedup is relatively small, especially when compared with already optimized rasterizers, Table 1 reports only 1.096× on Speedy-Splat. Compared with AdaGScale, the improvement is only 1.010× at standard resolution (Table 3). Also, much of the reported speedup over vanilla 3DGS comes from existing opacity-aware bounding techniques rather than TileSkipper itself. These results make me question whether the additional complexity is justified by the practical performance gains.

  2. Insufficient Justification for Regional Adaptation Table 2 shows that the proposed regional strategy achieves almost the same rendering performance as directly assigning thresholds to individual Gaussians. Regional adaptation improves over a global threshold, but it is unclear what additional benefits the proposed region grouping and allocation strategy provides compared with simpler per-Gaussian threshold selection.

  3. Missing Relevant Work The paper does not discuss FilterGS[1] (cvpr 2026), which also focuses on reducing redundant Gaussian-tile pairs for efficient rendering. Although the two methods target somewhat different settings, I believe this work should at least be discussed to better clarify the novelty of TileSkipper.

[1] FilterGS: Traversal-Free Parallel Filtering and Adaptive Shrinking for Large-Scale LoD 3D Gaussian Splatting

Questions for authors

My main questions concern the novelty over existing methods, the limited practical speedup, and the necessity of regional adaptation, as discussed in the Critical Weaknesses section.

Reviewer_1nVsOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TileSkipper, a method for speeding up 3D Gaussian Splatting by reducing the number of (Gaussian, tile) pairs processed during rendering. Standard renderers use the same cutoff for every Gaussian when determining which screen tiles it may affect. TileSkipper instead groups Gaussians and assigns different cutoffs based on estimates of how many Gaussian–tile pairs can be skipped and how much image error this may introduce. The cutoffs are selected separately for each scene using one set of training views and checked on another set of views. They are then reused for all subsequent camera views without modifying the trained 3DGS model. The experiments cover several 3DGS models and renderers, showing modest improvements at standard resolution and larger improvements at 4K.

Critical strengths
  • I appreciate that the authors evaluate the method with several 3DGS models and renderers. They also separate the improvemens due to existing techniques from those due to the proposed method. This makes the experimental evidence more credible.

  • Once the cutoffs have been selected, they can be applied directly during rendering without changing the trained 3DGS model or adding another computation step. There are also some empirical evidence that the same cutoffs can be be reused with different but compatible renderers.

Critical weaknesses
  • Technical novelty. Using a threshold to ignore weak Gaussian contributions is not new, as the authors acknowledge. Extending a single threshold to multiple thresholds for different Gaussians or regions is also a natural and straightforward idea, since different parts of a scene may tolerate different amounts of truncation. The main technical addition is therefore the procedure for choosing among a fixed set of cutoff values. This appears incremental. Moreover, grouping Gaussians into regions performs similarly to choosing a cutoff for each Gaussian separately and does not reduce storage, further weakening the main technical contribution.

  • Limited evidence of practical benefit. The speed improvement is modest at standard resolution and becomes more substantial mainly at 4K. Because the cutoffs must be selected separately for every scene, it is unclear whether the rendering time saved justifies the setup cost compared with a simpler global cutoff. The selected cutoffs are also fixed for all camera views, but the paper does not specifically test scenes with strong view-dependent appearance. Some unseen test views exceed the error limit used during selection.

Questions for authors
  • Table 6 reports setup times of 2.8 seconds for the scene-wide method and 9.0 seconds for the regional method. How many frames must be rendered before the additional setup cost of the regional method is recovered relative to the scene-wide method?
  • Do you have results on scenes with strongly view-dependent appearance or on camera views that are far from those used to select the cutoffs? These results would affect my assessment of the fixed-cutoff design.
  • How was the set of tested cutoff values chosen, and how sensitive are the results to this choice? For the optional pre-culling method, what safety inflation factor was used, and how was it selected?
Reviewer_MC63OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TileSkipper, a training-free, post-hoc optimization method for accelerating tiled 3D Gaussian Splatting (3DGS) rasterization. The central idea is to replace a scene-global contribution cutoff with a heterogeneous, per-Gaussian threshold policy that reduces the number of Gaussian–tile intersections while controlling rendering distortion.

The method estimates the computational savings and image distortion associated with candidate cutoff thresholds using calibration renders. It groups Gaussians into 64 regions, allocates threshold levels according to an estimated distortion budget, and validates the resulting policy using disjoint selection views. The final policy is stored as one byte per Gaussian and does not require retraining or per-frame inference.

The authors evaluate the method across 13 scenes from Mip-NeRF 360, Tanks & Temples, and Deep Blending, integrating it with multiple existing 3DGS models and rasterizers. They report a 1.088× average speedup at standard resolution and 1.238× at 4K in their fixed-policy AccuTile evaluation, with small average PSNR degradation. However, the incremental benefits over existing optimized rasterizers are substantially smaller, and the regional policy performs comparably to per-Gaussian threshold optimization.

Critical strengths

None.

Critical weaknesses
  1. Poor presentation and insufficient explanation of key concepts.

    • The paper is unnecessarily dense and difficult to follow. In particular, the introduction introduces method-specific notation and terminology, such as τ\tau, τ∗\tau^*, pair-count ratio (Figure 1), compiled threshold, and compiler, without adequately defining or explaining their meaning. These concepts are essential to understanding the proposed method, yet readers must consult later sections to interpret the introduction and its figures.
    • The motivation for several design choices is insufficiently explained. For example, it is unclear why Gaussians need to be grouped into regions, especially given that regional optimization performs comparably to per-Gaussian optimization (Table 2). Important details necessary for evaluating the contribution, including the attribution of speedups to prior optimizations, are relegated to later sections or the appendix, making it difficult to identify the actual improvements introduced by TileSkipper.
    • The conclusion (Section 5) largely repeats the method description without meaningfully summarizing the findings, discussing limitations, or explaining the significance of the proposed approach.
  2. Limited incremental speedup over existing optimized rasterizers.

    • For example, the 1.493× speedup on vanilla 3DGS consists of a 1.454× gain from the existing exact-bound optimization and only a 1.027× incremental gain from TileSkipper (Section 4.2).
  3. Limited methodological novelty.

    • The regional grouping strategy is not sufficiently justified. Table 2 shows that regional optimization performs nearly identically to per-Gaussian optimization, without reducing the stored policy size.
Questions for authors
  1. Motivation for regional grouping: What is the fundamental motivation for introducing region-based threshold allocation? Given that regional optimization performs nearly identically to per-Gaussian optimization (Table 2), what concrete advantage does this additional design choice provide?
  2. Motivation and necessity of the proposed method: Given the modest incremental speedups over existing opacity-aware rasterizers (e.g., 1.027× for vanilla 3DGS), what justifies the additional complexity of offline calibration and threshold optimization? Could a simpler heuristic achieve comparable results without the proposed compilation procedure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.