content_hash: d1b204ecfe8d9825

SharpGS: Sharpness-Preserving 3D Gaussian Splatting with Differentiable Blur-Driven Density Control¶

Conference: ECCV 2026
Paper: ECCV Official Portal
Area: 3D Vision
Keywords: 3D Gaussian Splatting, density control, differentiable blur, novel view synthesis, high-frequency detail recovery

TL;DR¶

Addressing the insensitivity of standard 3D Gaussian Splatting (3DGS) to intricate boundaries and high-frequency textures caused by gradient-only density control, SharpGS introduces blur as an explicit perceptual cue via a CUDA-accelerated differentiable blur framework, substantially elevating fine rendering quality while keeping the primitive representation compact through dual regularizations.

Background & Motivation¶

Novel view synthesis serves as a cornerstone for computer vision, virtual reality, and photorealistic real-time rendering. 3D Gaussian Splatting (3DGS) has rapidly emerged as a dominant alternative to Neural Radiance Fields (NeRFs), bypassing computationally demanding volumetric ray marching through explicit 3D Gaussian primitive representations paired with efficient tile-based 2D rasterization. Nonetheless, standard 3DGS heavily relies on view-space positional gradients to steer primitive densification (splitting and cloning). This gradient-driven heuristic exhibits prominent blind spots: in regions featuring intricate textures or fine object boundaries, positional gradients often saturate or cancel out across neighboring views, allowing oversized Gaussians to broadly over-cover detailed geometries and culminating in severe under-densification and fuzzy blur artifacts.

To alleviate over-reconstruction and fidelity loss, recent approaches have integrated frequency-domain regularizations (e.g., FreGS, Wavelet-GS), pixel-error-driven criteria (e.g., Revising-3DGS, ConeGS), or modified gradient metrics (e.g., AbsGS, Pixel-GS). However, frequency transformations commonly trigger substantial memory overhead and inflate primitive counts by more than $1.8\times$, whereas pixel-error signals tend to over-smooth high-frequency micro-textures. Furthermore, naively lowering the positional gradient threshold in 3DGS produces excessive Gaussian counts with negligible perceptual improvements, verifying that spatial gradient distributions fail to directly quantify the perceptual mismatch of high-frequency granularity between renderings and the scene.

This paper tackles the challenge from a compelling physical intuition: when 3DGS primitives are under-densified, the resulting visual degradation locally resembles a low-pass blurred version of the ground-truth (GT) image. If a GT image can be adaptively blurred to match the under-resolved reconstruction, the required degree of blur directly pinpoints regions where fine spatial details are lacking. Core idea: Introduce differentiable blur as an explicit perceptual cue to estimate per-primitive blur levels $\sigma$, coupled with detail-aware smooth background penalties and blur-adaptive soft opacity attenuation, enabling compact yet sharpness-preserving 3DGS density control.

Method¶

Overall Architecture¶

The SharpGS pipeline seamlessly incorporates a differentiable blur analysis module and two complementary regularization schemes into the standard 3DGS optimization loop. The workflow comprises three interacting stages: first, in the forward rendering pass, explicit 3D Gaussians are rasterized to synthesize the viewpoint image $I_{\text{gs}}$ alongside a pixel-space blur map $\sigma_{\text{pix}}$ aggregated from per-primitive blur parameters; second, the differentiable blur pipeline convolves the ground-truth image $I_{\text{gt}}$ with spatially varying kernels dictated by $\sigma_{\text{pix}}$ to generate a blurred counterpart $\hat{I}_{\text{gt}}$, and optimizes per-primitive blur attributes $\sigma_i$ under a joint detail-aware loss; finally, the densification pass relaxes the positional gradient threshold for primitives exhibiting high blur to selectively split fine structures, while periodically executing soft opacity attenuation and smooth-region regularization to suppress redundant primitives in structureless skies or entangled heterogeneous boundaries.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input viewpoint and 3D Gaussian primitive set"] --> B["1. Differentiable Blur Granularity Mismatch Perception<br/>Rasterize rendering and pixel blur map, convolve GT"]
    B --> C["2. Detail-Aware Regularization<br/>Penalize smooth background prior to prevent false positives"]
    C --> D["3. Blur-Adaptive Soft Opacity Attenuation<br/>Attenuate large overlapping primitives scaled by blur levels"]
    D --> E["Densification Criterion and Refinement<br/>Relax gradient split for high blur, prune low-opacity primitives"]
    E --> F["Output compact and sharpness-preserving 3DGS representation"]

Key Designs¶

1. Differentiable Blur Granularity Mismatch Perception: Mapping under-densification to adaptive blur kernels for pinpointing fine detail loss Standard 3DGS projects 3D Gaussians into 2D screen-space ellipses blended via alpha compositing. From a spatial sampling viewpoint, the color of an individual primitive can be interpreted as a gather-form weighted aggregation of local GT pixels within its projected footprint. Assuming neighboring primitives maintain comparable projected scales, the collective gathering effect across arbitrary screen pixels $p$ is approximated by an isotropic 2D Gaussian kernel $H_{\sigma_{\text{pix}}(p)}$: $$I_{\text{gs}}(p) \approx \hat{I}_{\text{gt}}(p) = \int_{u \in \Omega(\sigma_{\text{pix}}(p))} H_{\sigma_{\text{pix}}(p)}(u) I_{\text{gt}}(p-u) \, du$$ Building upon this equivalence, each 3D Gaussian is assigned a learnable scalar blur level $\sigma_i$ (parameterized via Sigmoid activation, scaled by $k_\sigma = 20$, and initialized to 0.3). A CUDA-based differentiable 2D Gaussian convolution pipeline (kernel radius $3\sigma_{\text{pix}}$) applies this spatially varying blur to $I_{\text{gt}}$. Backpropagation from the primary loss $L_p = (1 - \lambda_{\text{dssim}}) \|I_{\text{gs}} - \hat{I}_{\text{gt}}\|_1 + \lambda_{\text{dssim}} (1 - \text{SSIM}(I_{\text{gs}}, \hat{I}_{\text{gt}}))$ drives the blurred GT to align with the rendering. When high-frequency details are faithfully captured, $\sigma_{\text{pix}} \to 0$; conversely, when primitives under-represent local geometry, optimization pushes $\sigma_i$ higher. This parameter is decoupled from canonical appearance attributes and exclusively guides densification decisions.

2. Detail-Aware Regularization: Leveraging spatial structural priors to penalize smooth areas and prevent false positives Relying solely on matching the rendering with blurred GT easily triggers false positives in inherently textureless, smooth regions (e.g., clear sky or planar walls). In such zones, even strong blurring introduces negligible pixel discrepancies, allowing $\sigma_i$ to drift arbitrarily high without increasing $L_p$, which inadvertently triggers redundant primitive splits. To counteract this, a spatial detail-aware weight map $w_d$ is derived from the GT image by applying a fixed heavy blur (radius 60, corresponding to $\sigma_{\text{pix}} = 20$) and computing the L1 and SSIM discrepancy between the GT and its heavily blurred version. In smooth areas, $w_d \to 0$, whereas in high-frequency regions $w_d$ remains large. The detail-aware regularization loss is formulated as: $$L_{\text{reg}} = (1 - w_d) \frac{\sigma_{\text{pix}}}{k_\sigma}$$ Integrating this term yields the complete blur optimization objective $L_\sigma = (1 - \lambda_\sigma) L_p + \lambda_\sigma L_{\text{reg}}$ (with $\lambda_\sigma = 0.1$). $L_{\text{reg}}$ aggressively suppresses blur inflation across homogeneous backgrounds, eliminating spurious Gaussian proliferation.

3. Blur-Adaptive Soft Opacity Attenuation: Progressively subduing large redundant primitives based on blur magnitude While detail-aware regularization effectively controls homogeneous flat surfaces, heterogeneous regions (e.g., bicycle spokes intersecting distant background surfaces, or fine grass atop soil) present distinct challenges: large-scale primitives overlapping with intricate foreground details absorb elevated blur levels from surrounding high-frequency elements, leading to redundant densification. Standard 3DGS performs a coarse reset of all opacities near zero every 3000 iterations. SharpGS instead introduces soft opacity attenuation applied at a higher frequency (every 500 iterations), smoothly scaling the attenuation towards a baseline $\tau_m = 0.5$ in proportion to the normalized blur magnitude: $$o_{\text{new}} = \left(1 - \left(\frac{\sigma_i}{k_\sigma}\right)^\gamma\right) o_i + \left(\frac{\sigma_i}{k_\sigma}\right)^\gamma \tau_m$$ where $\gamma = 0.01$. Highly blurred, unreliable primitives suffer steeper opacity decay. Unless subsequent iterations reinforce their necessity through substantial photometric gradients, these primitives naturally diminish below the pruning threshold and are eliminated, preventing geometric clutter while preserving razor-sharp silhouettes.

Loss & Training¶

SharpGS keeps the standard 3DGS photometric objective $L_{\text{gs}} = (1 - \lambda_{\text{dssim}})\|I_{\text{gs}} - I_{\text{gt}}\|_1 + \lambda_{\text{dssim}}(1 - \text{SSIM}(I_{\text{gs}}, I_{\text{gt}}))$ ($\lambda_{\text{dssim}} = 0.2$) intact for updating Gaussian positions, covariances, colors, and opacities. The blur objective $L_\sigma$ exclusively propagates gradients to optimize the per-primitive blur parameters $\sigma$. During refinement iterations, a dual-criterion densification strategy is deployed: a primitive undergoes splitting or cloning if $\nabla_p L_{\text{gs}} \ge \tau_p$ or if its blur level meets $\sigma \ge \tau_\sigma$ under a relaxed positional gradient threshold $\nabla_p L_{\text{gs}} \ge \tau_p \cdot k_p$ (defaults: $\tau_\sigma = 2, k_p = 0.05$). Following densification, all $\sigma_i$ are reset to their baseline 0.3 to prevent cumulative bias across refinement cycles.

Key Experimental Results¶

Main Results¶

SharpGS was evaluated across three standard novel view synthesis benchmarks: Mip-NeRF360, Tanks&Temples, and Deep Blending. All models were trained for 30,000 iterations on a single NVIDIA A6000 GPU (48GB). The table compares canonical 3DGS, advanced variants, and their integration with SharpGS across rendering quality, primitive counts, and memory footprints.

Method / Dataset	Mip-NeRF360 PSNR ↑	Mip-NeRF360 SSIM ↑	Mip-NeRF360 LPIPS ↓	Mip-NeRF360 Primitives ↓	Tanks&Temples PSNR ↑	Tanks&Temples SSIM ↑	Tanks&Temples LPIPS ↓	Tanks&Temples Primitives ↓	Deep Blending PSNR ↑	Deep Blending SSIM ↑	Deep Blending LPIPS ↓	Deep Blending Primitives ↓
3DGS [16]	27.414	0.812	0.218	3350K	23.655	0.844	0.179	1893K	29.394	0.898	0.248	2833K
3DGS ($\tau_p=1.6\times 10^{-4}$)	27.541	0.818	0.205	4763K	23.772	0.848	0.169	2551K	29.176	0.896	0.247	3653K
Mip-Splatting [40]	27.557	0.816	0.217	3271K	23.963	0.852	0.176	1864K	29.598	0.904	0.243	2961K
Pixel-GS [45]	27.537	0.822	0.190	5622K	23.759	0.853	0.151	4598K	28.812	0.891	0.252	4623K
FDS-GS [42]	27.556	0.822	0.208	1542K	23.990	0.859	0.145	1530K	29.794	0.905	0.232	1711K
SteepGS [33]	26.999	0.792	0.249	2182K	23.393	0.838	0.193	1365K	29.418	0.903	0.252	1604K
Compact-GS [20]	26.989	0.796	0.245	1469K	23.360	0.831	0.199	928K	29.701	0.900	0.256	1142K
Ours + 3DGS	27.576	0.820	0.190	3083K	23.821	0.850	0.152	1658K	29.679	0.900	0.237	1836K
Ours + Mip-Splatting	27.679	0.824	0.199	2905K	24.047	0.856	0.162	1629K	29.802	0.906	0.237	1809K
Ours + Pixel-GS	27.652	0.822	0.182	4218K	23.900	0.855	0.138	2999K	29.620	0.897	0.235	2687K
Ours + FDS-GS	27.696	0.826	0.203	1501K	24.230	0.866	0.139	1499K	29.841	0.907	0.231	1491K
Ours + SteepGS	27.333	0.811	0.213	2112K	23.563	0.843	0.173	1206K	29.592	0.903	0.245	1315K
Ours + Compact-GS	27.070	0.805	0.225	1530K	23.560	0.836	0.189	906K	29.886	0.904	0.249	1028K

Ablation Study¶

A component-wise breakdown using 3DGS as the baseline confirms the respective roles of differentiable blur (Diff. blur), detail-aware regularization ($L_{\text{reg}}$), and blur-adaptive soft opacity attenuation (Opacity att.):

Configuration	Mip-NeRF360 PSNR/LPIPS/Prims	Tanks&Temples PSNR/LPIPS/Prims	Deep Blending PSNR/LPIPS/Prims	Empirical Insight
Baseline 3DGS [16]	27.414 / 0.218 / 3350K	23.655 / 0.179 / 1893K	29.394 / 0.248 / 2833K	Standard gradient-based baseline
+ Diff. blur	27.594 / 0.185 / 4213K	23.685 / 0.143 / 2586K	29.626 / 0.231 / 2565K	Perceptual blur cue sharply boosts visual details, but swells Gaussian counts
+ Diff. blur + $L_{\text{reg}}$	27.570 / 0.189 / 3621K	23.709 / 0.152 / 1977K	29.617 / 0.236 / 2140K	Smooth background penalty cuts 14%–23.6% redundant primitives with comparable quality
+ Diff. blur + Opacity att.	27.605 / 0.185 / 3520K	23.839 / 0.141 / 2061K	29.693 / 0.232 / 2197K	Soft attenuation effectively prunes broad overlapping primitives (16%–20.3% reduction)
FULL Model	27.576 / 0.190 / 3083K	23.821 / 0.152 / 1658K	29.679 / 0.237 / 1836K	Joint regularizations cut primitives by up to 35.2% vs. unregularized blur, surpassing baseline 3DGS

Key Findings¶

Substantial Perceptual Gains: SharpGS brings noticeable drops in LPIPS when integrated with 3DGS (from 0.218 down to 0.190 on Mip-NeRF360, and from 0.179 down to 0.152 on Tanks&Temples), corroborating that differentiable blur serves as an exceptionally sensitive perceptual indicator against blurring and muddy artifacts.
Superiority Over Gradient Brute-Force: Simply reducing the 3DGS gradient threshold to $\tau_p = 1.6 \times 10^{-4}$ expands the primitive budget to 4763K (+42%), yet only yields 27.541 dB PSNR. In contrast, Ours+3DGS delivers 27.576 dB with only 3083K primitives (8% fewer than baseline 3DGS), illustrating the efficiency of selective refinement.
Hyperparameter Robustness: Evaluating kernel radii $r \in \{30, 60, 90\}$ for the detail-aware weight map $w_d$ shows consistent PSNR performance on Mip-NeRF360 between 27.564 dB and 27.585 dB with primitive variations under 1.3%, showing that $w_d$ functions as a robust structural prior insensitive to tuning.

Highlights & Insights¶

Inverting Error Attribution via Differentiable Image Convolutions: Rather than attempting fragile, high-order gradient tracking or complex frequency decomposition to spot missing details, SharpGS creatively asks: "how much must the ground-truth image be convolved to match the current reconstruction?" Formulating this in pixel space via differentiable 2D Gaussian kernels provides smooth, noise-resilient supervisory gradients.
Plug-and-Play Synergy with Existing SOTA Baselines: The blur estimation pathway is entirely decoupled from standard appearance attributes. SharpGS uniformly enhances distinct architectures—including Pixel-GS, Mip-Splatting, and Compact-GS—improving visual sharpness while simultaneously pruning redundant elements (e.g., slashing Pixel-GS primitive counts on Mip-NeRF360 from 5622K to 4218K).
Negligible Training Overhead via Custom CUDA Kernels: Extending the CUDA rasterization pipeline with dedicated forward and backward passes for per-primitive blur adds a mere 2.52 ms per iteration at 1256×828 resolution, avoiding the crippling multi-hour training bottlenecks of deep perceptual metrics like LPIPS.

Limitations & Future Work¶

Resolution-Dependent Blur Scaling: Because the primitive blur $\sigma_i$ is optimized against screen-space pixel distances, significant deviations in target image resolution (e.g., at $\times 1.25$ scaling) allocate more primitives as identical structures cover larger pixel footprints. While a normalized metric $\tilde{\sigma}_i = \sigma_i / s$ alleviates this, fully autonomous multi-scale adaptation across heterogeneous camera sensors remains an open research path.
Susceptibility to Inherent Defocus or Motion Blur in Training Inputs: When source multi-view photographs inherently suffer from lens defocus or camera motion blur, the differentiable blur module may misattribute optical blur to structural under-densification, triggering unwarranted primitive splits in out-of-focus backgrounds. Coupling 3D deblurring frameworks with blur-driven densification presents a promising future direction.

vs. 3DGS [16] / AbsGS [37] / Pixel-GS [45]: Prior methods rely on spatial positional gradients which tend to saturate or cancel out across fine-grained textures and boundaries. SharpGS utilizes differentiable blur as an explicit perceptual cue alongside a relaxed gradient gate, guiding densification straight to under-represented geometries.
vs. FreGS [43] / Wavelet-GS [46]: Frequency-regularized methods rely on Fourier or wavelet transforms that induce massive primitive counts (often $1.8\times$ higher) and GPU memory crashes. SharpGS avoids global frequency transformations, achieving superior detail recovery in screen space with compact budgets.
vs. Revising-3DGS [28] / ConeGS [1]: Pixel-error densification frequently over-smooths delicate textures and ConeGS depends on precomputed iNGP depth priors; SharpGS captures granularity mismatches directly from image blur, preserving sharper high-frequency details with significantly reduced training times.

Rating¶

Novelty: ⭐⭐⭐⭐ [Framing under-densification as differentiable ground-truth blur and formulating paired regularizations is highly creative and elegant]
Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Thorough benchmarking across three major datasets, integration with six diverse 3DGS variants, and comprehensive ablation analyses]
Writing Quality: ⭐⭐⭐⭐⭐ [Clear motivation, solid mathematical foundations, intuitive narrative, and clean visual figures]
Value: ⭐⭐⭐⭐ [Lightweight, plug-and-play densification controller offering immediate practical value for high-fidelity novel view rendering]