Reflection-Sensitive Gaussian Pruning

• The paper introduces a gradient-based method that computes an importance score (I₍ᵢ₎) by blending base and reflection loss gradients to decide which Gaussian splats to prune.
• It achieves significant efficiency gains by adaptively removing low-impact splats, reducing model size nearly 4× and accelerating inference up to 7× without major quality losses.
• The strategy balances preserving photorealistic reflections with computational savings by integrating both gradient sensitivity and SDF-aware spatial criteria.

Reflection-sensitive Gaussian pruning refers to model reduction strategies in 3D Gaussian splatting pipelines where the decision to remove basis functions (Gaussians) is explicitly driven by their quantitative impact on both base and reflection image contributions. Such pruning is especially vital for realistic novel view synthesis in scenes featuring complex specular phenomena, where preserving splats that control reflections is essential for photorealism, while also accelerating inference and reducing memory. The two most prominent methods are the gradient-based reflection-sensitive pruning of HybridSplat and the SDF-based spatial pruning of GS-ROR$^2$.

1. Theoretical Formulation

Reflection-sensitive Gaussian pruning in HybridSplat (Liu et al., 9 Dec 2025) is fundamentally gradient-based. The model reconstructs two image branches per view: a “base” color $\mathbf{c}_{\text{base}}(\mathbf{p})$ and a “reflection” color $\mathbf{c}_{\text{ref}}(\mathbf{p})$, for each pixel $\mathbf{p}$. The final output is their convex combination: $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$ where $\beta$ is a global reflection blend factor.

The reconstruction loss is standard RGB $\ell_2$ loss over all pixels plus optional regularization. Each 3D Gaussian $g_i$ (base or reflective) is scored by the squared gradient magnitude of the loss with respect to its parameters: $$B_i = \|\nabla_{g_i} \mathcal{L}\|^2 \quad (g_i \in G_{\text{base}}), \qquad R_i = \|\nabla_{g_i} \mathcal{L}\|^2 \quad (g_i \in G_{\text{ref}})$$ The unified reflection-sensitive importance is: $I_i = (1-\beta) B_i + \beta R_i$ This scalar $\mathbf{c}_{\text{base}}(\mathbf{p})$0 quantifies the influence of $\mathbf{c}_{\text{base}}(\mathbf{p})$1 on the blended output, assigning reflection splats increased relevance where $\mathbf{c}_{\text{base}}(\mathbf{p})$2 is large. Gaussians with $\mathbf{c}_{\text{base}}(\mathbf{p})$3 (user-set threshold) or among the lowest $\mathbf{c}_{\text{base}}(\mathbf{p})$4 are pruned.

2. Reflection-Sensitive Pruning Algorithm

Pruning in HybridSplat operates either periodically during training or as post-processing. Each step computes the loss gradients via backpropagation, then evaluates $\mathbf{c}_{\text{base}}(\mathbf{p})$5 for each Gaussian. Pseudocode for adaptive threshold-based pruning:

$\ell_2$5

Aggressiveness is controlled by $\mathbf{c}_{\text{base}}(\mathbf{p})$6 or percentile $\mathbf{c}_{\text{base}}(\mathbf{p})$7; e.g., discarding 5% of splats every 500 iterations. The inclusion of the $\mathbf{c}_{\text{base}}(\mathbf{p})$8 term is mandatory to preserve highlight splats.

3. Integration into the Hybrid Splatting and Related Pipelines

HybridSplat schedules pruning after an initial warm-up—allowing base-reconstruction before culling. Once pruned, its tile-based data structures are rebuilt. The trained/pruned model then supports real-time inference with high reflection quality. Reflection-sensitive pruning ties directly to gradient feedback from both branches, ensuring that specular contributors are retained in proportion to their optical effect.

Conversely, GS-ROR$\mathbf{c}_{\text{base}}(\mathbf{p})$9 (Zhu et al., 2024) utilizes an SDF-aware pruning protocol complementary to deferred shading. After bidirectional supervision aligns Gaussian and SDF-predicted depths/normals, each Gaussian's central location $\mathbf{c}_{\text{ref}}(\mathbf{p})$0 is queried for its signed distance $\mathbf{c}_{\text{ref}}(\mathbf{p})$1. A masking function

$\mathbf{c}_{\text{ref}}(\mathbf{p})$2

with $\mathbf{c}_{\text{ref}}(\mathbf{p})$3 derived from a density falloff $\mathbf{c}_{\text{ref}}(\mathbf{p})$4, determines inclusion. Floaters ($\mathbf{c}_{\text{ref}}(\mathbf{p})$5) are pruned in each cycle, tightening as SDF sharpens, ensuring all retained Gaussians belong to the true surface.

4. Hyperparameter Selection and Trade-Offs

Reflection-sensitive pruning exposes trade-offs between model size, rendering speed, and scene reconstruction fidelity:

• Higher thresholds or larger prune fractions ($\mathbf{c}_{\text{ref}}(\mathbf{p})$6 or $\mathbf{c}_{\text{ref}}(\mathbf{p})$7) yield fewer splats, proportionally increasing speed and reducing memory usage. However, excessive pruning leads to loss of fine reflection features, visible as reduced highlight quality or “dull” reflections.
• Lower thresholds are more conservative, preserving image fidelity but limiting acceleration.
• Best practices: prune $\mathbf{c}_{\text{ref}}(\mathbf{p})$8 every 500 iterations and monitor held-out PSNR/SSIM, lowering $\mathbf{c}_{\text{ref}}(\mathbf{p})$9 if quality drops by more than $\mathbf{p}$0 dB. For real-time, $\mathbf{p}$1 can yield $\mathbf{p}$2 compression with tolerable (sub $\mathbf{p}$3 dB) PSNR loss.
• In scenes with subtle specularities, smaller $\mathbf{p}$4 is recommended to avoid under-representation of reflective details.

5. Quantitative Evaluation and Empirical Effects

HybridSplat’s reflection-sensitive pruning achieves substantial resource savings on complex reflective datasets. On Ref-NeRF and NeRF-Casting, measured against EnvGS (no pruning), the model size falls from approximately $\mathbf{p}$5 M to $\mathbf{p}$6 M Gaussians (almost $\mathbf{p}$7 reduction), and inference speed increases from $\mathbf{p}$8 FPS to $\mathbf{p}$9 FPS ($$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$0 speedup) on a single RTX4090. Quality metrics remain high: PSNR drops by less than $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$1 dB (EnvGS $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$2 dB, HybridSplat $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$3 dB), SSIM is $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$4 vs. $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$5, and LPIPS slightly increases, reflecting negligible visual degradation for major computational gain (Liu et al., 9 Dec 2025).

For GS-ROR$$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$6, introducing SDF-aware pruning after mutual supervision improves mean PSNR by $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$7 dB and SSIM by $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$8 (to $$\hat{I}(\mathbf{p}) = (1-\beta)\,\mathbf{c}_{\text{base}}(\mathbf{p}) + \beta\,\mathbf{c}_{\text{ref}}(\mathbf{p}),$$9/ $\beta$0), while removing floaters that degrade relighting quality. The overhead of SDF-based pruning is minimal, with real-time rendering ($\beta$1 FPS) and only a $\beta$2 h increase in total training time (Zhu et al., 2024).

System	Pruning approach	Splats (#, Ref-NeRF)	FPS (RTX4090)	PSNR/SSIM
EnvGS	No pruning	1.4M	15	30.21/0.872
HybridSplat	Reflection-sensitive	0.386M	107	29.87/0.864
GS-ROR$\beta$3	SDF-aware	(auto, scene dep.)	$\beta$4	23.31/0.9376

6. Practical Guidelines and Limitations

Reflection-sensitive pruning mandates the inclusion of the $\beta$5 term; omitting it preferentially removes splats providing key highlights. Best practice is to track validation PSNR, reducing prune rate if drops exceed $\beta$6 dB. In scenes dominated by diffuse content, $\beta$7 should be decreased to avoid over-pruning specular splats. Aggressive pruning (e.g., $\beta$8) is suitable only if mild degradation ($\beta$9 dB) is permissible.

In SDF-aware pruning as in GS-ROR$\ell_2$0, tying the mask threshold $\ell_2$1 to the dynamic shape of $\ell_2$2 ensures adaptivity: early in training, loose thresholds support geometric exploration, while later, tight thresholds guarantee surfacic fidelity.

A plausible implication is that reflection-sensitive pruning frameworks represent a general trend toward application-specific, task-aware model compression for neural rendering, where importance weighting directly tracks physically meaningful image contributions and not just generic error gradients.

7. Relationship to Broader Neural Rendering and Model Reduction Techniques

Reflection-sensitive pruning as pioneered in HybridSplat and GS-ROR$\ell_2$3 extends standard importance-based or spatial pruning to photorealistic, reflection-rich scene reconstruction, where naive methods would irrecoverably damage specular realism or introduce geometric floaters. The methodology integrates tightly into hybrid or deferred splatting pipelines and leverages physically rooted blend ratios, as well as signed distance functions for geometric regularization.

HybridSplat demonstrates that coupling importance scores across multi-branch rendering pipelines (by blending per-branch sensitivities) is critical for artifact-free, efficient novel view synthesis in challenging reflective environments. GS-ROR$\ell_2$4 shows mutual-supervision with SDFs both prunes geometric outliers and leads to sharper normals and specular highlights, all without runtime SDF dependence (Liu et al., 9 Dec 2025, Zhu et al., 2024).