Revising Densification in Gaussian Splatting

1 Introduction

• ADC’s positional-gradient thresholds can be unintuitive and miss high-frequency underfitting, while uncontrolled primitive growth can cause out-of-memory errors.The introduction also notes that ADC lacks explicit control of the maximum number of Gaussians per scene.Source: §1 · p. 2
• The revised approach reports consistent improvements over 3DGS and Mip-Splatting across benchmark datasets, with quantitative and qualitative gains.The evaluations include Mip-NeRF 360, Tanks and Temples, and Deep Blending.Source: §1 · p. 3
• The paper uses per-pixel error to guide densification, corrects opacity handling during cloning, and controls primitive counts per scene and densification run.Per-pixel errors are apportioned to contributing Gaussians in a camera-specific way; the maximum error per primitive is tracked across views and two ADC runs.Source: §1 · p. 3

1.1 Related works

• Mip-Splatting and other cited extensions retain the original ADC strategy and could potentially benefit from the revised approach.The authors report showing this for Mip-Splatting in their evaluation.Source: §1.1 · p. 4
• GaussianPro addresses related densification limitations through plane-guided primitive propagation, but focuses on planar regions rather than holistic densification.The authors say a fair standard Mip-NeRF 360 comparison was not feasible because GaussianPro’s improved SfM point cloud was not publicly shared at submission.Source: §1.1 · p. 4

2 Preliminaries: Gaussian Splatting

• Each Gaussian primitive has a center and covariance, plus opacity and a feature vector such as RGB color or spherical-harmonics coefficients.The primitive is represented as a 3D Gaussian kernel.Source: §2 · p. 4
• Splatting projects a 3D Gaussian into camera pixel space, approximating the projection to first order at the primitive’s center.The projected primitive is geometrically equivalent to a 2D Gaussian kernel; the projection Jacobian is evaluated at the center.Source: §2 · p. 4
• Rendering orders primitives by camera-space depth and combines their decoded features with alpha-compositing weights.With spherical-harmonics features, the decoder can produce RGB color; other decoders can render quantities such as depth or normals.Source: §2 · p. 5

3.1 Adaptive Density Control and its limitations

• ADC averages positional-gradient magnitudes across rendered views and grows a primitive when its score exceeds a user-defined threshold, splitting larger primitives and cloning smaller ones.The paper identifies unintuitive, model-sensitive thresholds, underfitting below threshold, and the lack of direct control over primitive counts as limitations.Source: §3.1 · pp. 5–6

3.2 Error-based densification

• Per-pixel errors are redistributed to primitives according to their contributions to rendered pixels, and the maximum per-primitive error across views between ADC runs prioritizes densification.The resulting score is expressed in terms of an error metric, making its threshold easier to set than a positional-gradient threshold.Source: §3.2 · pp. 6–7

3.3 Opacity correction after cloning

• Preserving opacity after cloning overweights cloned primitives; setting opacity to ˆα := 1 −√1 −α removes this bias in the simplified center-pixel case.Across all pixels, the correction reduces but cannot completely remove the bias; it is not applied to splits, whose opacity is preserved.Source: §3.3 · pp. 7–8

3.4 Primitives growth control

• A global primitive cap and per-run offspring budget constrain ADC growth, retaining the highest-scoring candidates when eligible primitives exceed the available budget.The example shows smooth growth to the allotted maximum without opacity-reset discontinuities; the control can also be applied to gradient-based densification.Source: §3.4 · pp. 8–9, Figure 3 · p. 9

3.5 Alternative to opacity reset

• Instead of periodically resetting all opacities, the method lowers each primitive’s opacity by 0.001 after each densification run, avoiding sudden changes in the densification metric while retaining sparsification.Because lowering opacity encourages use of the background and may create holes in novel views, the method also regularizes residual probabilities toward zero, weighted by 0.1.Source: §3.5 · pp. 9–10

4.1 Datasets and metrics

• The evaluation follows the 3DGS setup on Mip-NeRF 360, Tanks and Temples, and Deep Blending, measuring PSNR, SSIM, and LPIPS on every eighth image held out for validation.Mip-NeRF 360 includes nine scenes; the other datasets use Truck and Train, and Dr Johnson and Playroom, respectively.Source: §4.1 · p. 10

4.2 Experimental setup

• The experiments use re-implementations, keep ADC active for 27k iterations, and cap primitive counts at the corresponding baseline count for comparable models.Densification uses E_k > 0.1 and adds up to 5% of the current primitives per step.Source: §4.2 · p. 10
• The authors report corrected LPIPS values because they found that both public baseline codebases underestimated the metric; supplementary values enable comparison with the original tables.They say the discrepancy was confirmed in private correspondence with the authors.Source: §4.2 · pp. 10–11

4.3 Main results

• Across the Mip-NeRF 360, Tanks and Temples, and Deep Blending datasets, our improved ADC strategy consistently outperforms corresponding 3DGS and Mip-Splatting baselines, particularly on SSIM, over five training runs.Source: §4.3 · p. 11, Table 1 · p. 11, Table 2 · p. 11
• In qualitative comparisons at the same primitive count, the method ameliorates under-fitting artifacts while retaining 3DGS quality in non-problematic areas; Deep Blending shows a small, low-confidence PSNR regression.The authors associate remaining errors with flat, untextured surfaces and note that especially difficult scenes can still underfit.Source: Figure 4 · p. 12, §4.3 · p. 13

4.4 Ablation experiments

• On Mip-NeRF 360, the full method achieves the best results on all metrics; removing Opacity Correction, Growth Control, or Opacity Regularization degrades performance, with Opacity Correction having the largest impact.Growth Control slightly degrades results when removed from the full method, though the authors retain it as a capacity-control strategy.Source: Table 4 · p. 13, §4.4 · p. 14

4.5 Limitations

• Under-fitting can persist in especially difficult scenes, and perceptually improved regions may remain inaccurate against ground truth; the authors suspect limits involving view dependence, appearance variation, and splatting approximation.Examples include treehill, both Deep Blending scenes, and the flowers scene.Source: §4.5 · p. 14

• The paper addresses limitations in Adaptive Density Control for 3D Gaussian Splatting, a method for photorealistic rendering.Source: §5 · p. 14
• Its densification criterion uses per-pixel errors, alongside a mechanism to control the total number of primitives generated per scene.Source: §5 · p. 14
• The method corrects a bias in ADC opacity handling during cloning and reports consistent improvements over previous methods, particularly in LPIPS.Source: §5 · p. 14

A Finer-grained analysis of the MipNeRF360 results

• On flowers and treehill, LPIPS is better despite worse PSNR, and the renderings look visually better than the baseline.The authors note that LPIPS better correlates with perceptual similarity.Source: A Finer-grained analysis of the… · p. 17
• The paper presents additional qualitative results highlighting selected scenes from the Tanks and Temples and MipNeRF360 datasets in Fig. 5.Source: A Finer-grained analysis of the… · p. 17
• The method outperforms standard Gaussian Splatting on all MipNeRF360 scenes and metrics except PSNR on flowers and PSNR and SSIM on treehill.Scores are per-scene averages over five runs; the method's primitive cap is fixed to the baseline runs' median primitive count.Source: Table 5 · p. 17, A Finer-grained analysis of the… · p. 17

B Failure of standard Gaussian Splatting even with 10M primitives

• The experiment bypasses standard GS thresholding and uses the proposed growing strategy to reach 10M primitives.The comparison is on a particularly difficult Flowers validation view.Source: B Failure of standard Gaussian… · pp. 17–18
• At 10M primitives, the Flowers result remains substantially blurry, despite a slight increase in primitives in the critical area.The standard result uses 4.2M primitives; reaching 10M uses the proposed growing strategy with thresholding bypassed.Source: B Failure of standard Gaussian… · pp. 17–18
• The authors suggest underrepresented areas can score extremely low under gradient-based densification, preventing sufficient densification even at 10M primitives.Source: B Failure of standard Gaussian… · p. 18

C Ablation of different densification guiding error

• Despite ℓ1's weaker penalty for blurred areas, the method reconstructs grass that remains very blurry with standard Gaussian Splatting.The comparison is shown on a Flowers validation image.Source: C Ablation of different densification… · p. 18, Figure 7 · p. 21
• With ℓ1 guiding error, the method still outperforms Gaussian Splatting on all MipNeRF360 scenes and metrics, with a couple more exceptions than when using SSIM.Table 7 reports per-scene PSNR, SSIM, and LPIPS averages over five runs, with standard deviations.Source: C Ablation of different densification… · p. 18, Table 7 · p. 19
• Figure 5 compares ground truth, standard Gaussian Splatting, and the proposed method on highlighted Tanks and Temples and MipNeRF360 scenes.Source: Figure 5 · p. 20