MVGSR: Multi-View 3D Gaussian Splatting SR

• The paper introduces a framework that synergizes 3D Gaussian Splatting with multi-view consistency to super-resolve and enhance 3D scene details.
• It employs geometry-aware fusion, epipolar-constrained attention, and uncertainty modeling to prevent artifacts and ensure spatial coherence.
• The method demonstrates improved PSNR, SSIM, and LPIPS across benchmarks like Mip-NeRF360 and Tanks & Temples, validating its high-fidelity reconstructions.

Multi-View Consistent 3D Gaussian Splatting Super-Resolution (MVGSR) encompasses a family of frameworks and algorithms aiming to produce high-fidelity, high-resolution (HR) 3D reconstructions and novel view renderings from sets of low-resolution (LR) multi-view images using 3D Gaussian splatting (3DGS) representations. The core challenge is to super-resolve scene content while preserving cross-view geometric and textural consistency, thereby preventing artifacts and hallucinated details characteristic of naïve independently-applied 2D super-resolution techniques. MVGSR systems typically integrate explicit geometry-aware fusion, multi-view information, uncertainty management, and tailored loss functions to enforce both photometric quality and spatial coherence.

1. Fundamentals of 3D Gaussian Splatting and Super-Resolution

3D Gaussian Splatting forms the backbone of recent high-quality real-time novel view synthesis methods. A 3D scene is represented as a set of anisotropic Gaussian primitives

$$\mathcal{G} = \{g_i = (\mu_i, \Sigma_i, c_i, \alpha_i)\}_{i=1}^N$$

where $\mu_i \in \mathbb{R}^3$ is the center, $\Sigma_i \in \mathbb{R}^{3\times 3}$ the covariance (encoding scale and rotation), $c_i \in \mathbb{R}^3$ the color, and $\alpha_i$ the opacity. Rendering involves projecting each Gaussian onto the image plane as an elliptical kernel and compositing via alpha-blending along each camera ray. Differentiable volumetric splatting enables end-to-end optimization.

Super-resolution for 3DGS aims to enhance the spatial detail of the rendered novel views far beyond the native LR input images, but achieving this in a multi-view setup is fundamentally challenging. Independent 2D single-image super-resolution (SISR) introduces hallucinated per-view details, leading to multi-view inconsistency when the 3D representation attempts to explain conflicting pseudo-labels. MVGSR strategies explicitly address these conflicts by integrating geometric, photometric, and statistical cues for consistency.

2. Multi-View Guided and Geometry-Aware Fusion Strategies

One dominant class of MVGSR approaches leverages multi-view geometry and cross-view consistency in the synthesis of HR content. Strategies include:

• Auxiliary View Selection via Camera Poses: For each target view, informative auxiliary views are selected using geometric constraints (e.g., forward-looking direction and frustum overlap) and fused based on position/direction metrics. This reduces reliance on temporal continuity and accommodates arbitrary multi-view datasets (Zhang et al., 17 Dec 2025).
• Epipolar-Constrained Attention Mechanisms: Epipolar-guided spatial transformers restrict information aggregation to pixels lying on geometric epipolar lines, ensuring that content fused from auxiliary views is physically plausible and geometrically consistent with the target perspective (Zhang et al., 17 Dec 2025).
• Multi-View Voting Densification: Error maps between rendered and super-resolved pseudo-labels highlight underfitted regions. Back-projected 3D points—identified as inconsistent across multiple views—are used to drive localized densification of the Gaussian field only where necessary, avoiding redundancy (Xie et al., 24 May 2025).

These methodologies enforce that high-frequency textures and structures are only synthesized where multi-view evidence supports them, minimizing ghosting, popping, and floating artifacts.

3. Uncertainty Modeling and Loss Weighting

A central theme in MVGSR methods is the modeling of per-Gaussian or per-anchor uncertainty to modulate both supervisory signal and model growth:

• Variational Feature Learning: The HR-support feature vectors for each Gaussian or anchor are modeled as samples from a Gaussian distribution, parameterized by mean and log-variance, with the variance directly informing the confidence in local super-resolved detail (Xie et al., 2024, Xie et al., 24 May 2025).
• Uncertainty-Guided Supervision: Per-pixel uncertainty maps, rendered via the propagated variances, are used to weight the loss terms. High uncertainty suppresses the influence of potentially unreliable pseudo-labels, while reliable regions are reinforced during training (Xie et al., 2024).
• Density Control through Uncertainty: Gaussians or anchors exhibiting persistently high uncertainty are split, prompting finer localized modeling. Conversely, "floaters" (high-uncertainty, low-opacity primitives) are pruned to maintain compactness and fidelity (Xie et al., 2024).

This adaptive treatment ensures learning is focused on spatial regions where supervision is statistically trustworthy, significantly improving both convergence and output consistency.

4. Internal-External Knowledge Fusion and Selectivity

Recent MVGSR variants employ explicit mechanisms for balancing external image-level priors with internally derived 3DGS information:

• Mask-Guided Fusion: Cross-view inconsistency and domain gaps in external SISR priors are counteracted by per-pixel discrepancy masks that dictate, for each spatial location, whether to trust an external (2D SR or depth estimation) or internal (multi-scale 3DGS) source. This selective blending exploits the strengths of both, yielding sharper artifact-free reconstructions (Feng et al., 27 Nov 2025).
• Selective Super-Resolution via Fidelity Scores: Only scene regions insufficiently observed at high spatial fidelity in any LR view receive SISR-based supervision, while all others rely on real high-frequency LR content. Gaussian-wise fidelity metrics guide selective injection of super-resolved signals, maximally preserving multi-view consistency (Asthana et al., 1 Dec 2025).

Such strategies reduce overfitting to hallucinated SISR content and prevent propagation of view-dependent artifacts, outperforming naive uniform SR fusion approaches.

5. Optimization Pipelines and Loss Formulations

MVGSR frameworks universally adopt two-stage or alternating optimization pipelines, with significant algorithmic details including:

• Coarse-to-Fine Training: An initial LR latent scene representation is learned (often via multi-resolution hash grids), capturing global structure and geometry. This is followed by a fine HR stage, focused on detail enrichment, densification, and uncertainty-aware refinement while the coarse model is kept frozen or minimally updated (Xie et al., 2024, Xie et al., 24 May 2025).
• Pseudo-Label Construction: SISR networks (e.g., SwinIR) and depth SR models produce HR pseudo-labels for photometric and geometric supervision. MVGSR employs strategies to mitigate error accumulation from these external networks, such as multi-view joint learning (Xie et al., 2024).
• Weighted/Masked Losses: Reconstruction losses (L₁, SSIM, intra-view photometric) are weighted per-pixel by uncertainty or selectivity masks. Regularization on uncertainty, volume, and perceptual similarity (LPIPS) is commonly applied (Xie et al., 2024, Feng et al., 27 Nov 2025, Asthana et al., 1 Dec 2025).
• Explicit Multi-View Consistency Losses: Joint optimization over multiple sampled target views, with summed or fused gradients, and explicit averaging or voting schemes for error-driven densification and supervision (Xie et al., 24 May 2025, Xie et al., 2024).

The balanced integration of these algorithmic elements yields models that simultaneously maximize resolution, consistency, and representational compactness.

6. Quantitative and Qualitative Outcomes

Empirical evaluations consistently show that MVGSR approaches surpass prior state-of-the-art baselines in both traditional reference-based (PSNR, SSIM, LPIPS) and no-reference/consistency metrics (e.g., FID, cross-view error):

Method	Dataset	PSNR ↑	SSIM ↑	LPIPS ↓	Key Finding
SuperGS (Xie et al., 2024)	Mip-NeRF360 (×4)	27.12	0.768	0.262	Outperforms prior by +0.24 dB
IE-SRGS (Feng et al., 27 Nov 2025)	Mip-NeRF360 (×4)	27.15	0.779	0.278	Further boosts SSIM and LPIPS
SplatSuRe (Asthana et al., 1 Dec 2025)	Tanks & Temples	23.81	0.784	0.272	Maximal gains in difficult areas
MVGSR (Zhang et al., 17 Dec 2025)	NeRF-Synthetic (×4)	33.01	0.9655	0.0368	Best overall fidelity, consistency

In qualitative analysis, MVGSR methods recover textural and geometric detail that prior uniform-SR methods cannot, suppressing multi-view artifacts and yielding sharper, artifact-free reconstructions in both synthetic and real-world settings.

7. Limitations and Prospective Directions

Notable limitations of current MVGSR methods include:

• Dependence on External Priors: Heavy SISR and depth SR models employed for pseudo-label generation slow training and may introduce domain-specific hallucinations or inconsistencies (Xie et al., 2024).
• Per-Scene Optimization: Most frameworks require re-training per scene and lack generalizable pre-trained variants (Xie et al., 2024).
• Threshold and Hyperparameter Sensitivity: Performance is sensitive to mask thresholds, uncertainty thresholds, and densification/voting parameters (Xie et al., 24 May 2025, Feng et al., 27 Nov 2025).
• Monocular Depth Limitation: Use of monocular depth priors restricts geometric fidelity; multi-view or depth fusion priors are anticipated to further advance geometry (Feng et al., 27 Nov 2025).

Future work is focused on the development of generalizable, scene-agnostic SR models, joint end-to-end optimization of 2D priors and 3DGS, improved occlusion reasoning via learned or adaptive geometric sampling, and extension to dynamic or temporally varying scenes using spatio-temporal consistency signals (Xie et al., 2024, Feng et al., 27 Nov 2025, Zhang et al., 17 Dec 2025).