Technical Summary: CLEAR - Conflict-aware Learning via Evidence-guided Adaptive Routing for Unified Sparse-View 3D Gaussian Super-Resolution

Problem Statement

Sparse-view 3D Gaussian Splatting (3DGS) Super-Resolution (SR) aims to reconstruct high-quality, high-resolution (HR) 3D scenes from sparse and low-resolution (LR) input views. This task is inherently challenging due to the lack of sufficient geometric and high-frequency information in the inputs.

Existing approaches typically follow a two-stage pipeline:

1. Reconstructing an LR Gaussian field from sparse inputs.
2. Transferring this field to an HR representation and refining it.

This paradigm suffers from several limitations:

• Error Accumulation: Errors from the initial LR reconstruction cascade into the HR refinement stage.
• Pipeline Complexity: The separation prevents end-to-end optimization.
• Gradient Conflicts: When attempting to unify LR and HR supervision in a single stage (as proposed by this paper), conflicting gradients arise. Specifically, as the number of input views decreases, the proportion of Gaussians where LR and HR gradients point in opposing directions (gradient conflict) increases significantly, leading to severe novel-view artifacts.

Methodology: CLEAR Framework

The authors propose CLEAR (Conflict-aware Learning via Evidence-guided Adaptive Routing), the first unified single-stage framework for sparse-view 3D Gaussian super-resolution. Instead of separate LR/HR representations, CLEAR optimizes a single unified Gaussian field throughout training, jointly leveraging authentic LR observations and external HR priors.

The framework consists of three core components:

1. Gaussian-Wise Conflict-Aware Optimization (GCO)

To address the gradient conflicts between authentic LR supervision and external HR priors, CLEAR implements a conflict-aware strategy at the individual Gaussian level.

• Conflict Detection: For each Gaussian $i$, the cosine similarity ($c_i$) between the LR gradient ($g^{LR}_i$) and the HR gradient ($g^{HR}_i$) is computed. A conflict is identified when $c_i < -\tau_c$ (where $\tau_c$ is a tolerance threshold).
• Soft Correction: Rather than discarding conflicting HR gradients entirely (which may contain useful detail cues), the method applies an evidence-conditioned soft correction. The LR gradient serves as a reliable anchor. If a conflict is detected, the HR gradient is adjusted to align with the LR direction, weighted by an adaptive correction factor. This suppresses destructive updates while preserving beneficial cross-resolution corrections.

2. Evidence-Guided Patch-to-Gaussian Routing (P2G)

To recover high-frequency details without introducing artifacts, CLEAR introduces a routing mechanism that selectively injects trustworthy details.

• Patch Reliability & Demand: The method estimates two metrics for each HR patch:
  • Reliability ($R_p$): Measures consistency between the downsampled HR image and the original LR image.
  • Detail Demand ($A_p$): Measures the high-frequency content in the HR patch.
• Lifting to Gaussian Space: These patch-level metrics are projected onto the visible Gaussians to generate a joint evidence score ($w_i$).
• Selective Routing:
  • Gradient Routing: The joint evidence modulates the HR detail gradients, ensuring high-frequency learning occurs primarily where the SR evidence is reliable.
  • Densification Routing: The evidence guides Gaussian densification (splitting/cloning), ensuring new Gaussians are added only in regions requiring detail recovery, thereby suppressing unreliable SR artifacts.

3. Sparse-View Stabilization (SVS)

To mitigate overfitting and geometric drift common in sparse-view optimization, CLEAR employs two stabilization strategies:

• Shared Gaussian Dropout: A shared dropout mask is applied to both LR and HR rendering scales during training. This prevents artificial inconsistencies between the two scales and reduces overfitting.
• Detached Mid-Training Anchoring: Once the Gaussian structure stabilizes at a specific iteration ($t_a$), densification is halted. A "detached anchor" (a stop-gradient copy of the state at $t_a$) is introduced. A regularization loss is applied to keep the current parameters close to this anchor, suppressing late-stage geometric drift while allowing limited appearance refinement.

Key Contributions

1. Unified Single-Stage Framework: CLEAR is the first method to perform sparse-view 3D Gaussian super-resolution in a unified single-stage optimization, avoiding the error accumulation and complexity of two-stage pipelines.
2. Conflict-Aware Optimization: The paper identifies and quantifies the severity of LR-HR gradient conflicts under sparse views. It proposes a Gaussian-wise strategy that uses LR gradients as anchors to suppress destructive HR updates while preserving compatible corrections.
3. Evidence-Guided Routing: A novel mechanism that lifts patch-level SR reliability and detail demand into the Gaussian field, selectively guiding high-frequency learning and densification to achieve reliable detail reconstruction.
4. Stabilization Strategy: The introduction of shared Gaussian dropout and detached mid-training anchoring enhances geometric consistency and generalization in sparse-view settings.

Experimental Results

CLEAR was evaluated on three 4× super-resolution benchmarks: LLFF (real-world, 3 views), Blender (synthetic, 8 views), and Mip-NeRF 360 (real-world, 24 views).

• Quantitative Performance: CLEAR consistently achieved State-of-the-Art (SOTA) results across all datasets, outperforming existing 3DGS-based methods (e.g., SRGS, SplatSuRe, DropGaussian) and two-stage baselines (S2Gaussian).
  • On LLFF (3 views), CLEAR achieved a PSNR of 20.42 and LPIPS of 0.266, surpassing the previous best (S2Gaussian) by a significant margin.
  • On Blender and Mip-NeRF 360, CLEAR also secured the highest PSNR and SSIM scores and the lowest LPIPS scores.
• Qualitative Performance: Visual comparisons demonstrated that CLEAR preserves better scene geometry and reconstructs finer textural details (e.g., foliage, railings) with significantly fewer artifacts compared to prior methods.
• Efficiency: CLEAR demonstrated superior training efficiency, requiring significantly less time than the two-stage S2Gaussian baseline (612s vs. 1028s on average for LLFF).

Significance and Claims

The paper claims that CLEAR effectively reconciles the inherent tension between structural cues from sparse LR observations and high-frequency information from external HR priors. By unifying these objectives within a single Gaussian field and employing conflict-aware and evidence-guided mechanisms, CLEAR achieves:

• Superior Rendering Quality: Higher fidelity in both geometry and appearance.
• Geometric Fidelity: Reduced artifacts and drift in novel views.
• Efficiency: A streamlined single-stage optimization process that is faster than multi-stage alternatives.

The authors position CLEAR as a robust solution for applications requiring high-quality 3D reconstruction from limited, low-resolution data, such as virtual/augmented reality, robotics, and digital twins.