acceptodds
Under review as a conference paper at ICLR 2027

VDGS: Improving Fidelity in Fast 3D Gaussian Splatting Training via Virtual Densification

Abstract

Fast 3D Gaussian Splatting training must improve fidelity with limited densification. Existing candidate criteria often aggregate gradients or errors at the parent level. Visible-view averages do not directly capture cross-view frequency, while parent aggregation obscures demand across the sides of a prescribed split. We propose VDGS, based on Virtual Densification (VD), to rank candidate splits before creating children. Parent gradient statistics first nominate candidates; the prescribed split geometry then partitions their pixelwise backward signals into two virtual child domains. VDGS combines two-sided activity and net-gradient direction differences, accumulates this evidence over a selection window, and ranks splits for submission. It reuses the regular backward pass without temporary children or trial training. In a common low-budget framework, full VDGS improves mean PSNR over direct AbsGS growth by approximately 0.45 dB on nine Mip-NeRF 360 scenes, while also improving SSIM and LPIPS.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Adaptive Densification, Efficient Training, Differentiable Rendering, Novel View Synthesis

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

This paper introduces Virtual Densification to evaluate potential splits before creating child Gaussians. It partitions gradients into two virtual child regions and combines gradient activity and directional differences to rank split candidates. The proposed method also accumulates scores across training views to guide densification. The method achieves more efficient splitting with certain Gaussian primitive budget. Experiments show competitive reconstruction quality and training efficiency compared with existing fast 3DGS methods.

Critical strengths

The proposed Virtual Densification provides an intuitive way to evaluate split potential using pixelwise gradients. This method does not explicitly creating child Gaussians or performing additional optimization, which is computationally efficient. The experiments show better performance compared with baselines.

Critical weaknesses

Table 4 shows that Accumulated Parent ABS already achieves strong performance. Accumulated HC provides only marginal improvements in SSIM and LPIPS, while slightly decreasing PSNR. This suggests that much of the improvement comes from window accumulation and the overall densification pipeline rather than the proposed two-sided scoring mechanism. This raises concerns about the effectiveness of the proposed method.

Questions for authors

Please refer to "Critical Weaknesses"

Reviewer_RrjFOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VDGS. VDGS is a method that improves fast 3DGS by selecting more useful splits under a limited densification budget. It virtually evaluates each prescribed split using the parent Gaussian’s existing pixelwise gradients, checking whether both prospective child regions have strong but different optimization demands. These scores are accumulated across sampled views to guide splitting, without creating temporary children or running trial training. Experiments show improved reconstruction quality at competitive training times.

Critical strengths

The paper addresses a practical problem: allocating limited densification capacity effectively during fast 3D Gaussian Splatting training. Its split-aligned evaluation has a clear intuition and reuses existing backward-pass gradients. Unlike some other 3DGS densification methods, it avoids temporary children or trial training. It requires less computation resources than most baseline methods.

Critical weaknesses

The comparisons provided in this paper may be insufficient. In addition, the provided results in this paper might be sub-optimal. Some baselines are missing. I just list an example below.

e.g., there's a paper in ICML 2026 introducing novel densification strategy by comparing patchwise error: PCGS: Deblurring 3D Gaussian Splatting with Patch Comparison https://openreview.net/pdf?id=4bLAhR3KK7

I copied results of PCGS, and compare them with VDGS here:

  • Mip-NeRF360 PSNR: PCGS 28.01, VDGS 27.95
  • T&T PSNR: PCGS 24.40, VDGS 24.39
  • Deep Blending PSNR: PCGS 30.04, VDGS 30.32

It seems that VDGS doesn't have a very large advantage. While I agree the tasks of the two works are somehow different, they share the common roadmap of novel densification technique, so the authors should provide experiments against PCGS to validate your superiority, or at least give some discussions, e.g., is PCGS significantly slower than your method?

In addition, I wonder if the virtual evaluation of prescribed splits (bottom part of fig.1) will make the method consume significantly more GPU memory. Could you please provide some relative VRAM usage info?

Questions for authors

My major concern is about the missing baselines, see Critical Weaknesses


EDIT: after rebuttal the rating is updated from 2 to 3

Reviewer_RTHhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Gaussian Splatting represents a scene using Gaussian elements. Splitting these into smaller elements can capture more details but this increases training cost. This paper proposes VDGS to help choose which splits are worth making. For each candidate split, VDGS groups the existing Gaussian’s pixel contributions into 2 regions corresponding to the proposed children. It uses training gradients to measure how strongly each region needs adjustment and whether their adjustment directions differ. The scores are accumulated across training views to select splits without creating temporary Gaussians or running trial training. The proposed approach was evaluated on Mip-NeRF360, Tanks and Temples, and Deep Blending. The complete pipeline improves quality over FastGS at comparable training times.

Critical strengths
  1. The proposed complete method improves reconstruction quality while keeping training fast. On Mip-NeRF360, Tanks and Temples, and Deep Blending, VDGS improves all three reported quality metrics over FastGS at comparable training times. Comparisons with matched Gaussian budgets and growth schedules also show improvements.

  2. The virtual probe reuses existing training gradients to assess proposed splits. It groups an existing Gaussian’s pixel contributions into 2 regions corresponding to a possible split. It then checks the gradient strength and direction in each region. The scores are collected across training views before selecting which Gaussians to split. This allows proposed splits to be examined without the cost of creating and training temporary Gaussians.

Critical weaknesses
  1. The paper’s main idea is to check the gradients on both sides of a possible Gaussian split before deciding whether to create new Gaussians. However, Table 4 shows that simply adding up the gradient activity of each whole Gaussian across training views already gives most of the quality improvement. This simpler method does not examine the two sides separately. The proposed two-sided score gives slightly lower PSNR and small improvements in the other quality metrics. Full VDGS also adds rules to prioritise regions that need more details and to fill unused split slots. However, the simpler method is not tested with these same rules. As a result, it is not clear where the gains come from.

  2. The paper does not show how VDGS performs compared to DC4GS, which is a previous method that also examines gradients in separate regions of a Gaussian to decide where to split it. Table 3 compares only DC4GS’s rule for selecting which Gaussians to split. It leaves out the procedure that examines those regions and chooses the split location. Without comparing against this component under a similar training budget, the added value of VDGS in reconstruction quality or training time is not demonstrated.

Questions for authors
  1. Could you compare full VDGS with accumulating each whole Gaussian’s gradient activity across training views? Give both methods the same rules for prioritising Gaussians with high gradient activity and filling unused split slots.

  2. Could you compare VDGS with DC4GS when including its rule for selecting Gaussians and its procedure for choosing where to split them? Evaluate how reconstruction quality and training time compare under similar training budgets.

  3. How sensitive are the results to the distance between the proposed child Gaussians, the number of iterations used to accumulate scores, and how often the virtual probe runs?

Comments on the paper

1 comment

We thank the reviewers and Area Chair for the discussion. The VDGS revision retains the original virtual densification method and adds controls for the two-sided score, comparisons with related methods, and measurements of sensitivity and memory. The main changes are marked in blue.

  1. Contribution of the two-sided score (pp. 9 and 16, Tables 4 and 10). We added a nine-scene comparison with accumulated Parent ABS using the same demand protection, deficit filling, and accumulation schedule. HC reduces LPIPS in all nine scenes, by 0.0035 on average, and increases mean PSNR by 0.02 dB. The revision credits window accumulation and the overall pipeline for much of the gain. Table 4 now gives 27.65 dB for the full-VDGS reference fixed before running the protected Parent ABS control. Table 2 retains the original overall result of 27.67 dB.
  2. DC4GS and the two-window pipeline (pp. 8–9 and 16). The shared-pipeline comparison now includes both DC4GS parent selection and adaptive split placement. Under matched budgets, full VDGS improves mean PSNR by 0.47 dB and reduces LPIPS by 0.0355 relative to DCC + DCS, with mean training times of 83.90 s and 89.03 s, respectively. Table 5 brings the all-parent versus two-window comparison into the main text. Two-window scoring reduces the mean peak number of probed parents by 81.6%, with similar training time.
  3. Sensitivity and GPU memory (p. 10, Tables 6–8). We report sensitivity to the split coefficient, accumulation window, and probe frequency. Full VDGS has lower three-scene mean LPIPS at every tested setting. Its peak PyTorch allocation is 21.84 MiB (1.23%) above visible-mean Parent ABS on average.
  4. PCGS and implementation details (pp. 15 and 17). Appendix Table 12 places published PCGS and VDGS results alongside their respective hardware and point budgets. Appendix A.5 specifies the inherited rotation and spherical-harmonic coefficients and the initialization of child opacity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.