SplatSuRe: Selective Super-Resolution for Multi-view Consistent 3D Gaussian Splatting (2512.02172v1)

Abstract: 3D Gaussian Splatting (3DGS) enables high-quality novel view synthesis, motivating interest in generating higher-resolution renders than those available during training. A natural strategy is to apply super-resolution (SR) to low-resolution (LR) input views, but independently enhancing each image introduces multi-view inconsistencies, leading to blurry renders. Prior methods attempt to mitigate these inconsistencies through learned neural components, temporally consistent video priors, or joint optimization on LR and SR views, but all uniformly apply SR across every image. In contrast, our key insight is that close-up LR views may contain high-frequency information for regions also captured in more distant views, and that we can use the camera pose relative to scene geometry to inform where to add SR content. Building from this insight, we propose SplatSuRe, a method that selectively applies SR content only in undersampled regions lacking high-frequency supervision, yielding sharper and more consistent results. Across Tanks & Temples, Deep Blending and Mip-NeRF 360, our approach surpasses baselines in both fidelity and perceptual quality. Notably, our gains are most significant in localized foreground regions where higher detail is desired.

• The paper introduces SplatSuRe, a framework that selectively injects super-resolution based on a per-Gaussian fidelity metric to maintain multi-view consistency.
• It employs a geometric-driven SR weighting and a composite loss that blends low-resolution supervision with spatially weighted SR losses, reducing multi-view artifacts.
• Experimental evaluations on multiple datasets show improved perceptual quality and detail recovery, outperforming uniform super-resolution approaches.

SplatSuRe: Selective Super-Resolution for Multi-view Consistent 3D Gaussian Splatting

Introduction and Motivation

The SplatSuRe framework addresses a persistent challenge in high-fidelity novel view synthesis with 3D Gaussian Splatting (3DGS): the inability to reliably reconstruct high-resolution (HR) details from low-resolution (LR) supervision, especially when naive super-resolution (SR) techniques break multi-view consistency. Traditional methods, such as single-image SR or uniform SR application during training, frequently inject view-inconsistent high-frequency content, leading to spatial artifacts and loss of coherence in multiview 3D reconstructions. SplatSuRe departs from prior art by introducing a geometric, view-aware selectivity to the injection of SR supervision, aiming to realize sharper renderings while maintaining strict cross-view consistency.

Figure 1: Overview of the SplatSuRe framework for selective super-resolution in multi-view 3D Gaussian Splatting.

Problem Formulation and Prior Art

Contemporary 3DGS approaches deliver real-time photorealism but are fundamentally bounded by the spectral content present in their training images. Existing strategies for SR in 3DGS, such as SRGS and SuperGaussian, apply super-resolved guidance uniformly or rely on learned regularizers to reduce inconsistency. However, they ignore inherent disparities in the density and proximity of LR multi-view coverage—i.e., close LR views can already densely sample certain regions, obviating the need for additional generated details there. This indiscriminate SR application gives rise to conflicts in the optimization objective, particularly in poorly or multiply-constrained regions observable from different camera perspectives.

Figure 2: Disparity in ground-truth frequency content; proximate LR views sample sufficient detail, obviating SR in many regions, but necessitating it in others.

Methodology

Geometric-Driven, Adaptive SR Weighting

Central to SplatSuRe is the computation of a per-Gaussian fidelity metric: a ratio of maximal to minimal projected 2D radii across all views per Gaussian primitive, quantifying the maximal spatial frequency sampled in each region by the available LR perspective images. A Gaussian with a high radius ratio indicates substantial variance in sampling density across views, suggesting cross-supervision exists; a ratio near unity signals uniform, potentially under-resolved supervision, motivating SR guidance. A smoothly thresholded sigmoid function maps these ratios to fidelity weights.

Super-resolution region selection then proceeds by rendering per-pixel weight maps per view using these fidelity scores. Notably, pixels associated with Gaussians most closely observed in a given view, or which lack any high-frequency constraints from other views, are assigned high SR weights. The 3DGS model is then trained using both an LR ground-truth loss (on downsampled renders) and a spatially weighted SR loss on the HR render, modulated by these per-view, per-pixel weights.

Figure 3: Super-resolution weight maps—bright: undersampled, SR-needed regions; dark: areas well-resolved by other views.

Loss Function Formulation

The combined loss, $$\mathcal{L} = (1-\gamma)\mathcal{L}_{LR} + \gamma\mathcal{L}_{SR}$$, seamlessly integrates consistent LR supervision and selective, geometry-aware SR enhancement. This design prevents superfluous generative detail in over-constrained regions and limits multi-view artifacts typical of isotropic SR guidance.

Experimental Evaluation

The method is comprehensively validated on Tanks & Temples, Deep Blending, and Mip-NeRF 360 datasets, with $4\times$ and $8\times$ SR. Across all metrics (PSNR, SSIM, LPIPS, FID, DreamSim, CMMD), SplatSuRe exhibits consistent or superior quantitative performance, particularly in scenarios with high disparity in camera-to-object sampling and substantial LR downsampling.

Figure 4: Qualitative results show SplatSuRe produces sharper, more coherent renderings compared to Mip-Splatting and SRGS baselines.

There is a pronounced improvement in textural sharpness, the recovery of fine structures, and reduction in SR-induced artifacts. The method’s advantage is most evident in scenes where input LR views are heterogeneous in their coverage—foreground and sparsely observed objects benefit most.

Further analysis investigates the impact of the SR selection ratio threshold. Excessively aggressive SR (large thresholds) introduces inconsistencies, while soft selectivity enables optimal perceptual and pixel-wise fidelity.

Figure 5: Effect of SR selection (ratio threshold) on rendering quality; excessive SR degrades performance.

Ablation and Sensitivity Studies

Ablations show the generality of the approach for various SR backbones (SwinIR, StableSR) and across geometric scene variations. The method is robust to the underlying SISR model, consistently outperforming uniform-SR approaches, and demonstrates that improvements stem primarily from selective injection rather than backbone choice.

Implications and Future Directions

The proposed selectivity mechanism sets a precedent for geometry- and sampling-aware data fusion in generative 3D reconstruction. Its compatibility with both generative and deterministic SR models allows seamless integration into existing pipelines, offering a plug-in solution for fidelity enhancement. While SplatSuRe primarily addresses color and detail fidelity, extending the paradigm to multi-scale or uncertainty-adaptive weighting could offer further advantages. Incorporation of scene semantics or content-aware refinement could further improve SR guidance, particularly in boundaries or regions of complex geometry. The approach also naturally synergizes with view-consistent diffusion-based 3D priors, and may facilitate higher fidelity in mesh or NeRF-based 3D reconstructions.

Conclusion

SplatSuRe introduces an explicit, geometric framework for selective super-resolution in 3DGS, leveraging a principled fidelity metric to guide SR where it is most necessary. This strategy yields improved perceptual and quantitative reconstruction quality, particularly in challenging undersampled regions, while suppressing multi-view inconsistencies. The paradigm of geometrically-aware, region-selective enhancement establishes a new direction for robust, high-resolution neural rendering and invites extensions into broader 3D generative modeling applications.