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

Source: arXiv:2608.02206 · Published 2026-08-03 · By Hantang Li, Qiang Zhu, Xiandong Meng, Debin Zhao, Xiaopeng Fan

TL;DR

This paper addresses the challenging task of sparse-view 3D Gaussian Splatting super-resolution, where the input observations are both sparse and low-resolution, lacking sufficient geometric and high-frequency information for high-quality 3D reconstruction. Existing methods typically use a two-stage pipeline: first reconstructing a low-resolution Gaussian field and then refining it to high resolution, which causes error accumulation and inefficiencies. CLEAR is proposed as the first unified single-stage framework that jointly optimizes a single Gaussian field constrained simultaneously by authentic low-resolution views and externally generated high-resolution priors. This unification avoids stage-wise transfer errors but introduces gradient conflicts between low- and high-resolution supervision. To handle these conflicts, the authors develop a Gaussian-wise conflict-aware optimization strategy that treats low-resolution gradients as reliable anchors, softly correcting high-resolution gradients only when they strongly conflict. They also introduce an evidence-guided Patch-to-Gaussian routing mechanism which estimates patch-level reliability and detail demand to selectively propagate trustworthy high-frequency information into the Gaussian space, helping to recover fine details. Additional training stabilization techniques, including shared Gaussian dropout and a detached mid-training anchor, improve robustness. Experiments on multiple synthetic and real-world 4× super-resolution benchmarks demonstrate CLEAR consistently achieves state-of-the-art rendering quality and superior geometric fidelity compared to prior two-stage and multi-stage approaches, while also training faster. Ablation studies confirm the importance of each novel component in reducing gradient conflicts and improving reconstruction. Overall, CLEAR effectively reconciles sparse-view structural cues with high-frequency detail guidance within a single-stage unified 3D Gaussian optimization framework.

Key findings

• CLEAR achieves PSNR of 20.42 on LLFF 4× super-resolution with 3 input views, outperforming the previous best two-stage method S2Gaussian (20.17) and other multi-stage baselines.
• On Blender dataset with 8 input views, CLEAR achieves PSNR 24.26 and LPIPS 0.087, improving over S2Gaussian's 23.78 PSNR and 0.104 LPIPS.
• On Mip-NeRF 360 with 24 views, CLEAR delivers PSNR 22.19, SSIM 0.697, and LPIPS 0.288, surpassing S2Gaussian results (PSNR 21.96, SSIM 0.669).
• Gaussian-wise conflict-aware optimization reduces the average LR-HR gradient conflict ratio from 0.243 (baseline) to 0.228, while improving PSNR from 18.74 to 19.72 on LLFF.
• Evidence-guided Patch-to-Gaussian routing further reduces conflict ratio to 0.210 and improves PSNR to 20.09.
• Sparse-view stabilization achieves final PSNR 20.42 and conflict ratio 0.162, the lowest among tested variants.
• Training time on LLFF with 3 views is approximately 612 seconds for CLEAR vs 1028 seconds for two-stage S2Gaussian, showing efficiency.
• Qualitative results show CLEAR better preserves geometric structures and recovers high-frequency details under sparse views across LLFF, Blender, and Mip-NeRF 360 datasets.

Threat model

The implicit adversary in this context is the optimization conflict arising from jointly training on authentic sparse low-resolution observations and externally generated high-resolution priors, where these two sources provide conflicting gradients that could degrade final reconstruction quality. The authors assume the LR views are reliable anchors, and the HR priors are informative but potentially conflicting. The optimization framework cannot rely on perfect alignment and must mitigate destructive updates due to such gradient conflicts. The adversary cannot corrupt input images or priors but represents the intrinsic difficulty in combining heterogeneous supervision under sparse data.

Methodology — deep read

The paper tackles sparse-view 3D Gaussian Splatting super-resolution, where input cameras are few and low-resolution, and an external high-resolution image prior is used. The threat here is effectively optimizing a Gaussian scene representation under conflicting LR and HR supervision.

Data: Experiments use LLFF (8 scenes with 3 views), Blender synthetic (8 views), and Mip-NeRF 360 (24 views). Low-resolution images are inputs; high-resolution references are generated via a frozen image SR model (ResShift). Metrics are PSNR, SSIM, LPIPS.

Architecture: CLEAR maintains a single unified Gaussian field Gθ representing the scene explicitly as anisotropic 3D Gaussians with trainable parameters (positions, covariance, appearance). It renders both LR and HR views from the same Gaussian set using scale-aware Mip rasterization. The Gaussian appearance splits into a base shared feature and a detachment HR residual.

Loss: Jointly optimizes LR photometric+SSIM loss (authentic observation) and HR losses from external SR references, consisting of base photometric+SSIM and a Haar-domain high-frequency loss.

Optimization challenge: LR and HR gradients for each Gaussian can conflict strongly, especially under sparse views where LR supervision is weak but trustworthy, and HR priors can introduce erroneous updates.

Conflict-aware optimization: For each Gaussian i, compute cosine similarity between LR gradient gLR_i and HR gradient gHR_i. When similarity falls below a threshold (τ_c=0.05), apply a soft correction to the HR gradient based on the severity of conflict and patch-level reliability evidence, ensuring that LR gradients anchor the optimization and destructive HR updates are suppressed.

Patch-to-Gaussian routing: Evaluate each HR image patch for reliability (based on LR-HR consistency using downsampled HR) and detail demand (high-frequency content). Lift this patch-level evidence into the Gaussian field, weighting the HR gradients and densification of Gaussians accordingly. This selective routing focuses high-frequency learning where reliable evidence exists.

Sparse-view stabilization: To mitigate overfitting and geometric drift due to limited views, apply shared Gaussian dropout masks consistently across LR and HR rendering during training, and introduce a detached mid-training anchor regularization once the geometry stabilizes, preserving structure while allowing residual appearance refinement.

Training: Use Adam optimizer on a single NVIDIA A800 GPU for 10K iterations. LR patches of 16×16 and corresponding HR patches of 64×64 are used. Loss weights and thresholds are set empirically.

Inference directly renders the final Gaussian field at target HR resolution without stage-wise transfer or refinement.

Evaluation: Quantitative metrics evaluated on held-out views across LLFF, Blender, and Mip-NeRF 360 scenes, compared to NeRF- and 3DGS-based baselines including two-stage variants like S2Gaussian. Ablations progressively add conflict-aware optimization, patch evidence routing, and stabilization to confirm their contributions. Training times are compared to two-stage methods.

A concrete example end-to-end: Given sparse LR input views and their parameters, CLEAR generates HR references via ResShift. During training iteration, the Gaussian field is rendered in LR and HR; gradients computed from LR and HR losses for each Gaussian. Gaussian-wise conflict is detected; HR gradients are corrected softly based on conflict severity and patch evidence weights. Corrections are applied and combined LR+HR gradients update Gaussian params. Shared dropout masks regularize. Past anchor state used after geometry stabilizes. Iterations continue until 10K; resulting model renders crisp HR novel views with improved fidelity.

Technical innovations

• A unified single-stage framework that jointly optimizes a single Gaussian field under both LR authentic views and external HR priors, avoiding error accumulation in two-stage pipelines.
• Gaussian-wise conflict-aware optimization strategy that treats LR gradients as reliable anchors and applies evidence-conditioned soft correction only to severe HR gradient conflicts.
• Evidence-guided Patch-to-Gaussian routing mechanism that estimates local patch reliability and detail demand, lifting these into Gaussian space to selectively route high-frequency gradients and Gaussian densification.
• Sparse-view stabilization using shared Gaussian dropout during multi-scale rendering and detached mid-training Gaussian anchoring to reduce overfitting and geometric drift.

Datasets

• LLFF — 8 scenes with sparse views (3 views in evaluation) — Real-world dataset
• Blender — Synthetic scenes with 8 views — Standard synthetic dataset
• Mip-NeRF 360 — 360-degree scenes with 24 views — Real-world dataset

Baselines vs proposed

• S2Gaussian (two-stage baseline): PSNR=20.17, SSIM=0.640, LPIPS=0.314 on LLFF (3 views) vs CLEAR: PSNR=20.42, SSIM=0.682, LPIPS=0.266
• S2Gaussian: PSNR=23.78, SSIM=0.876, LPIPS=0.104 on Blender (8 views) vs CLEAR: PSNR=24.26, SSIM=0.880, LPIPS=0.087
• S2Gaussian: PSNR=21.96, SSIM=0.669, LPIPS=0.298 on Mip-NeRF 360 (24 views) vs CLEAR: PSNR=22.19, SSIM=0.697, LPIPS=0.288
• Baseline w/o proposed methods: PSNR=18.74, Conflict ratio=0.243 on LLFF vs full CLEAR: PSNR=20.42, Conflict=0.162
• DropGaussian + SRGS: PSNR=19.14, LPIPS=0.326 on LLFF vs CLEAR: PSNR=20.42, LPIPS=0.266

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2608.02206.

Fig 1: Motivation and Performance. (a) Unlike existing two-stage methods, CLEAR is the first unified single-stage

Fig 2: Overview of the proposed CLEAR framework. Given sparse LR observations, CLEAR optimizes a Gaussian field

Fig 3 (page 2).

Fig 4 (page 2).

Fig 5 (page 2).

Fig 6 (page 2).

Fig 7 (page 2).

Fig 8 (page 2).

Limitations

• The approach depends on external high-resolution priors generated by pretrained 2D image super-resolution models, potentially limiting applicability where such priors are not available or reliable.
• Evaluation is limited to 4× super-resolution and relatively small input view counts (3, 8, 24), leaving open the method's scalability to more extreme sparse or lower input resolutions.
• No explicit adversarial or worst-case robustness evaluation against corrupted or misleading HR priors is reported.
• The Gaussian-wise conflict threshold τ_c and other hyperparameters are empirically set and may require tuning for new datasets or configurations.
• The approach focuses on Gaussian splatting; results may not generalize to other 3D representations like explicit meshes or voxel grids without adaptation.
• No publicly released code or pretrained models at time of writing, which may slow community adoption and reproducibility.

Open questions / follow-ons

• How would CLEAR perform under varying types or qualities of high-resolution priors, including noisy or biased 2D super-resolution inputs?
• Can the conflict-aware optimization be extended to other 3D scene representations beyond Gaussian splatting, such as neural implicit fields or mesh-based models?
• What is the behavior and performance of CLEAR under more extreme sparse views, fewer than 3 or partial occlusions?
• Could adversarial or robust training methods further enhance reliability against misleading HR priors or corrupted observations?

Why it matters for bot defense

While CLEAR focuses on 3D Gaussian super-resolution from sparse views rather than traditional CAPTCHA or pure bot-defense domains, its core technical contribution—conflict-aware training to fuse reliable but sparse authentic observations with external higher-resolution priors—offers lessons for unified optimization under heterogeneous and partially conflicting supervision. For bot-defense and CAPTCHA systems, learning to reconcile reliable anchor signals with stronger but potentially conflicting auxiliary evidence could inspire robust classification or verification architectures that withstand adversarial or low-data regimes. The evidence-guided routing mechanism illustrates a data-driven selective fusion approach that weighting cues by contextual reliability and detail relevance, potentially valuable in multi-modal bot detection models that integrate disparate signals. Furthermore, the stabilization techniques to combat overfitting and geometric drift under sparse supervision highlight generalizable principles for improving robustness in security-critical ML pipelines. Practitioners might explore analogous conflict-aware updates or patch-level evidence weighting to fuse multiple weak and strong signals in CAPTCHA-solvers or bot classifiers, particularly when high-fidelity labels are scarce or noisy.

Cite

@article{arxiv2608_02206,
  title={ CLEAR: Conflict-aware Learning via Evidence-guided Adaptive Routing for Unified Sparse-View 3D Gaussian Super-Resolution },
  author={ Hantang Li and Qiang Zhu and Xiandong Meng and Debin Zhao and Xiaopeng Fan },
  journal={arXiv preprint arXiv:2608.02206},
  year={ 2026 },
  url={https://arxiv.org/abs/2608.02206}
}

Read the full paper

• arXiv:2608.02206 (abstract)
• arXiv:2608.02206 (PDF)