Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

SuperGS: Consistent and Detailed 3D Super-Resolution Scene Reconstruction via Gaussian Splatting

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read SuperGS claims a new state of the art in high-resolution novel view synthesis by combining a coarse-to-fine feature field, error-based multi-view voting densification, and uncertainty-guided supervision.

desk verdict SuperGS is a coherent engineering contribution with consistent reported gains, but the SOTA claim rests on an uncertainty mechanism that is not validated against a controlled pseudo-label baseline. read the letter →

arxiv 2505.18649 v1 pith:LUL4ZP2X submitted 2025-05-24 cs.CV

classification cs.CV
keywords 3DGaussiansplattinghigh-resolutionnovelviewsynthesisscenesuper-resolutionmulti-viewconsistencydensificationuncertaintymodelingvariationalfeaturespseudo-labelsupervision
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that 3D Gaussian Splatting, which normally produces coarse renderings when trained on low-resolution views, can be upgraded to high-resolution novel view synthesis by combining a coarse-to-fine latent feature field with a densification rule driven by multi-view agreement. The proposed method, SuperGS, first learns a low-resolution scene representation whose hash-grid feature field is frozen and reused as the foundation for a second high-resolution stage, where 2D single-image super-resolution outputs serve as pseudo-labels. To keep those per-view pseudo-labels from corrupting the 3D model, SuperGS replaces gradient-based Gaussian growth with error-based depth backprojection plus multi-view voting, and it models per-anchor feature uncertainty as a variational distribution whose variance down-weights unreliable supervision. The paper reports that this combination outperforms prior high-resolution novel view synthesis methods on Mip-NeRF360, Deep Blending, Tanks&Temples, and LLFF, with perceptual metrics (FID) improving by as much as 41.7% over a strong baseline.

What carries the argument

The load-bearing mechanism is the replacement of Scaffold-GS's gradient-based anchor growth with a multi-view voting densification. Candidate anchor positions come from backprojecting pixels whose rendered-versus-pseudo-label loss exceeds a threshold: $X_{k,i}=R_k^T K_k^{-1}(u_{k,i} d_{k,i}, v_{k,i} d_{k,i}, d_{k,i})^T + t_k$, where $d_{k,i}$ is the depth at that pixel from a high-resolution depth map obtained by rendering low-resolution depth and upsampling it with a guided depth super-resolution model. Candidates cast votes into a multi-resolution voxel grid, and new anchors are added in voxels whose vote count exceeds a threshold, at the appropriate granularity. The other half of the machinery is variational anchor features: the supplementary feature is $f_{\text{sr}}^{\text{supp}} = f_\mu + \epsilon \exp(f_\sigma)$ with $\epsilon \sim \mathcal{N}(0,1)$, so the per-anchor uncertainty $u_i = \|\exp(f_\sigma)\|_2$ can be $\alpha$-blended into an uncertainty map that reweights the reconstruction loss and triggers anchor refinement by splitting uncertain anchors into smaller voxels.

What would settle it

On a synthetic scene with known geometry, replace the guided depth super-resolution model with bilinear upsampling of the rendered low-resolution depth; if the voting densification still improves PSNR over gradient-based growth, the central depth-accuracy premise is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the inconsistencies of 2D image super-resolution pseudo-labels are not a fatal obstacle for 3D super-resolution, provided the optimization is made consistency-aware in three specific ways. First, a low-resolution latent feature field, built on a multi-resolution hash grid with contracted coordinates, is kept frozen through the fine stage so high-resolution detail grows on top of globally consistent structure. Second, densification is driven by the pseudo-label error map: pixels with large loss are backprojected into 3D through high-resolution depth maps (rendered low-resolution depth upsampled by a guided depth super-resolution model), and new anchors are created only where multi-resolution voxels collect enough votes from multiple views. Third, each anchor's supplementary feature is sampled from a Gaussian distribution via the reparameterization trick, and the variance magnitude acts as an uncertainty signal that both splits uncertain anchors into finer voxels and reweights the reconstruction loss pointwise. With these components, the paper claims state-of-the-art results on both forward-facing and 360-degree high-resolution novel view synthesis benchmarks.

Load-bearing premise

The densification assumes the upsampled high-resolution depth maps are geometrically accurate enough that backprojected error pixels land on genuinely under-reconstructed surface points; a biased or noisy depth prior would place new anchors in wrong 3D locations, degrading the reconstruction.

Editorial extensions

If this is right

  • On the reported ablations, the multi-view voting densification improves PSNR by 0.54 dB over gradient-based growth on LLFF while using fewer than half as many Gaussians (693,781 vs 1,723,549).
  • The uncertainty-guided supervision directly reduces the influence of pseudo-label regions the model has low confidence in, which should make the method more robust to SISR artifacts and hallucinated textures.
  • Because the low-resolution feature field is frozen during fine training, the method inherits a coherent scene structure, implying that the coarse-to-fine design itself is a defense against per-view 2D prior drift.
  • Reported FID gains (e.g., 34.65 on LLFF, a 41.7% improvement over Mip-splatting) indicate the method produces images closer in distribution to ground-truth high-resolution photos, not just higher pixel-wise fidelity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's error-based voting densification is not tied to SISR outputs; the same vote/backprojection machinery could be driven by diffusion-based or video-based super-resolution priors, and the paper's experiments do not test this extension.
  • The rendered uncertainty map could be exposed as a per-view confidence output useful for downstream editing or safe-use decisions, rather than only a training-time weight, which the paper does not discuss.
  • A sharper test of the depth assumption would ablate the guided depth upsampler by substituting bilinear upsampling; if gains persist, the method's dependence on geometric accuracy is weaker than claimed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces SuperGS, a Scaffold-GS-based method for high-resolution novel view synthesis (HRNVS) from low-resolution inputs. It uses a two-stage coarse-to-fine framework: a hash-grid latent feature field is trained on low-resolution views, then a high-resolution stage adds SISR pseudo-labels, multi-view voting densification based on backprojected upsampled depth maps, and variational anchor features whose predicted uncertainty reweights the reconstruction loss and triggers anchor splitting. The authors report consistent improvements over 3DGS, Scaffold-GS, Mip-Splatting, SRGS, and cited baselines on Mip-NeRF360, Deep Blending, Tanks&Temples, and LLFF, with gains in PSNR, SSIM, LPIPS, and FID.

Significance. If the results hold, SuperGS is a practical contribution: it shows a way to combine 2D SISR priors with 3D Gaussian representation while addressing multi-view consistency through geometry-guided densification and uncertainty weighting, and it evaluates on multiple real-world datasets with held-out ground truth, avoiding circularity in the benchmark. The main strengths are the clear problem formulation, the use of standard metrics, and the explicit attempt to reduce Gaussian redundancy. However, the current evidence for the headline claim is weakened by single-run evaluations, ablation only on LLFF, and a few load-bearing design choices that lack validation; those issues are addressable but require additional experiments.

major comments (5)
  1. [Section 4.3.3, Eq. (15)] The uncertainty-weighted loss L_rec = ||(1-sigma(U)) ⊙ (I_pred - I_pseudo)||_1, with U rendered from anchor uncertainties u_i = ||exp(f_sigma)||_2 (Eq. 11), creates a trivial loss-reduction path: for any pixel where the pseudo-label is hard to fit, the model can increase f_sigma, lowering (1-sigma(U)) and hence the gradient from that pixel, without fitting the detail. Because f_sigma is learned end-to-end against the same pseudo-labels, the reported PSNR/FID gains may partly reflect suppression of inconsistent labels rather than true HR reconstruction. A controlled test is needed, e.g., supervise the same framework with clean ground-truth HR images and verify the uncertainty mechanism still helps, or show that uncertainty is higher in regions where pseudo-labels are known to be inconsistent.
  2. [Section 4.2, Eq. (10)] The densification stage backprojects pixels with high error using high-resolution depth maps produced by rendering low-resolution depth (Eq. 9) and upsampling with a GDSR model. The entire candidate-position generation therefore inherits any bias or noise in the low-resolution depth estimate; the paper provides no validation against ground-truth geometry or against geometric consistency across views. Without such a check, the claimed advantage of multi-view voting over gradient-based densification could be due to details of the voting threshold rather than to correct geometry. Please add a quantitative analysis (e.g., depth error vs. GT on a dataset where it exists) or an ablation replacing the depth prior with a monocular depth oracle.
  3. [Section 5, Tables 1-3] All quantitative results are single-run and no variance or seed information is reported, and the ablations in Table 3 are only on the 9-scene LLFF set. The main additive gains are small (0.08 dB for uncertainty, 0.54 dB for densification); single-run differences of this size are not convincing evidence of improvement. Please report mean ± std over multiple seeds or at least per-scene results and run the ablations on at least one 360-degree dataset.
  4. [Section 5.1.2 and Table 2] NeRF-SR is listed as a baseline by citing its published numbers, but the Related Work (Section 2.2) states that NeRF-SR requires ground-truth high-resolution images as input. If so, the reported comparison is not under identical conditions: SuperGS is trained with SISR pseudo-labels, whereas NeRF-SR uses oracle HR supervision, and the +0.30 dB claim in Section 5.2.1 is therefore not an apples-to-apples gain. Either re-run NeRF-SR under the same pseudo-label protocol or clearly state the comparison is against a method with a different supervision regime.
  5. [Section 5.1.3 and Sections 4.2.2/4.3.2] Several thresholds and weights that control the core contributions are either unspecified or listed once without sensitivity analysis: the multi-view voting threshold, the error-map threshold ("loss exceeding 0.08"), the uncertainty refinement threshold, the softmax temperature beta in Eq. (9), and the loss weights lambda_SSIM, lambda_vol, lambda_LPIPS in Eq. (16). Without a sensitivity study, the reader cannot tell whether the reported improvements are robust or are tuned to the test sets. Please provide the full configuration and an ablation or sweep for at least the voting and uncertainty thresholds.
minor comments (5)
  1. [Eq. (10)] The backprojection omits the camera-center offset under the standard world-to-camera extrinsic convention (X = R^T(K^{-1}[u d, v d, d] - t)); please state the convention used for [R_k, t_k].
  2. [Section 4.1.2] The text says "we model f_sr_field as a variational feature" but the distribution is defined on f_sr_supp; please unify the notation.
  3. [Tables 1 and 2] Table 2 has no LPIPS or FID entries for NeRF-SR, and Table 1 has missing GaussianSR entries for Tanks&Temples; please explain or fill in the missing values.
  4. [Figure 4] The scene and training configuration for the reported Gaussian counts (693,781 vs 1,723,549) are not given; please specify them.
  5. [Section 4.1.1] The text cites the Mip-NeRF360 contraction approach as reference [1], but reference [1] is Mip-NeRF; please correct the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported gains are measured against held-out high-resolution ground truth, and training-time pseudo-labels are not themselves the target quantity.

full rationale

SuperGS's central claim is high-resolution novel-view quality, and the reported numbers in Tables 1 and 2 are computed against genuine held-out high-resolution captures (Mip-NeRF360, Deep Blending, Tanks&Temples, LLFF), not against the pseudo-labels used in training. The SISR pseudo-labels enter training (Sec. 4.3.3, Eq. 15) and the GDSR depth map powers densification (Sec. 4.2), but no test quantity is defined in terms of these priors. The uncertainty term in Eq. 15 could in principle reduce loss by down-weighting inconsistent pseudo-label regions, but that is a training-dynamics concern, not a circular reduction: the model cannot make PSNR/FID against real images high by construction. The only author-overlap citation ([49], listed as an example of 3DGS adaptation for spectacular scenes) is background and not load-bearing; no uniqueness theorem or ansatz is imported from prior work by the same authors. The derivation chain is therefore self-contained against external ground truth, and no circular step is exhibited.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several unvalidated domain assumptions about the reliability of 2D priors (SISR and GDSR) and the meaningfulness of learned uncertainty. The free parameters are training hyperparameters, not derived constants; they are typical for ML methods but their values are largely unspecified, which affects reproducibility.

free parameters (6)
  • Error-map loss threshold = 0.08
    Pixels with reconstruction loss exceeding 0.08 are selected for depth backprojection (Section 4.2.2, 5.1.3). The value is chosen by hand and affects where new anchors are proposed.
  • Multi-view voting threshold = not specified
    Anchors are added to voxels when the count of candidate points exceeds a threshold (Section 4.2.2). The threshold is not reported; it is a free parameter tuned by the authors.
  • Uncertainty refinement threshold = not specified
    Anchors with uncertainty exceeding a threshold are split (Section 4.3.2). Threshold value is not reported.
  • Softmax temperature beta = not specified
    Beta in Eq. 9 controls depth weight differentiation; value not reported.
  • Loss weights lambda_SSIM, lambda_vol, lambda_LPIPS = not specified
    Weights in Eq. 16 are not reported; likely tuned for the datasets.
  • Hash grid levels L = not specified
    Number of resolution levels in the latent feature field (Eq. 5) is not reported; affects representation capacity.
assumptions (6)
  • domain assumption SISR pseudo-labels (SwinIR) are reliable enough to guide high-resolution training
    The fine stage uses SwinIR upsampled images as supervision (Section 4.1.2). If pseudo-labels contain large systematic errors, the uncertainty weighting may not compensate.
  • domain assumption GDSR-upsampled depth maps of rendered low-res depth are geometrically accurate
    Section 4.2.1 uses a GDSR model to upsample rendered depth; Section 4.2.2 backprojects these depths to 3D. No depth accuracy verification is provided.
  • domain assumption Variational feature variance correlates with pseudo-label inconsistency
    The paper assumes uncertainty (Eq. 11) identifies unreliable regions (Section 4.3.1), but does not validate this correlation directly.
  • domain assumption Latent feature field from coarse stage is a good foundation for high-res stage
    Freezing the feature field (Section 4.1.2) assumes low-res features transfer to high-res detail; ablation shows only 0.15 dB gain.
  • domain assumption Multi-view voting in voxels proxies multi-view consistency
    The voting mechanism (Section 4.2.2) counts candidate points across views, assuming agreement among views indicates reliable regions; no theoretical or empirical justification beyond improved metrics.
  • standard math Reparameterization trick yields meaningful uncertainty
    Eq. 8 uses the standard reparameterization trick for variational features; this is a standard mathematical tool.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SuperGS: Consistent and Detailed 3D Super-Resolution Scene Reconstruction via Gaussian Splatting." pith.science (2026). https://pith.science/paper/LUL4ZP2X

@misc{pith2026250518649,
  author       = {Pith},
  title        = {Pith review of: SuperGS: Consistent and Detailed 3D Super-Resolution Scene Reconstruction via Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LUL4ZP2X}},
  note         = {Machine review of arXiv:2505.18649}
}
read the original abstract

Recently, 3D Gaussian Splatting (3DGS) has excelled in novel view synthesis (NVS) with its real-time rendering capabilities and superior quality. However, it encounters challenges for high-resolution novel view synthesis (HRNVS) due to the coarse nature of primitives derived from low-resolution input views. To address this issue, we propose SuperGS, an expansion of Scaffold-GS designed with a two-stage coarse-to-fine training framework. In the low-resolution stage, we introduce a latent feature field to represent the low-resolution scene, which serves as both the initialization and foundational information for super-resolution optimization. In the high-resolution stage, we propose a multi-view consistent densification strategy that backprojects high-resolution depth maps based on error maps and employs a multi-view voting mechanism, mitigating ambiguities caused by multi-view inconsistencies in the pseudo labels provided by 2D prior models while avoiding Gaussian redundancy. Furthermore, we model uncertainty through variational feature learning and use it to guide further scene representation refinement and adjust the supervisory effect of pseudo-labels, ensuring consistent and detailed scene reconstruction. Extensive experiments demonstrate that SuperGS outperforms state-of-the-art HRNVS methods on both forward-facing and 360-degree datasets.

Figures

Figures reproduced from arXiv: 2505.18649 by the authors.

Figure 1
Figure 1. Comparisons in High-resolution Novel View Synthesis. SuperGS demonstrates more detailed and high-fidelity results compared to existing GS-based methods, delivering consistent advantages in extremely challenging scenes. Abstract Recently, 3D Gaussian Splatting (3DGS) has excelled in novel view synthesis (NVS) with its real-time rendering capabilities and supe￾rior quality. However, it encounters challenges for high-r… view at source ↗
Figure 2
Figure 2. Framework of our proposed SuperGS. (a) We propose a two-stage coarse-to-fine framework. We enhance Scaffold-GS by introducing a latent feature field to represent the low-resolution scene, which serves as both initialization and foundational information for super-resolution optimization. (b) In the high-resolution stage, we propose a multi-view consistent densification strategy that replaces the original gradient-bas… view at source ↗
Figure 3
Figure 3. Illustration of Feature Field. For a specific anchor, we extract and interpolate features from hash tables using its coordinates, with the concatenation of features from 𝐿 resolution levels forming the field feature 𝑓field of this anchor. Within this approach, 𝑓field is derived from a latent feature field, which learns low-resolution scene information and serves as the foundational feature for subsequent high-resolu… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of Ours Densification and Gradient￾based Densification Startegy. Our method achieves better reconstruction quality with fewer anchor points, reducing memory requirements while preventing overfitting. 4.2.1 Depth map rendering and super-resolution. Depth maps…
Figure 5
Figure 5. Figure 5: Illustration of Densification Strategy. We introduce a multi-view voting densification strategy that replaces the original anchor growing policy. First, we generate candidate Gaussian positions by back-projecting the corresponding pixels with depth map based on pixel-w…
Figure 6
Figure 6. Figure 6: Qualitative comparison of the HRNVS (×4) on real-world datasets. We highlight the difference with colored patches. 3DGS Scaffold-GS Mip-splatting SRGS Ours Ground Truth [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of the HRNVS (×4) on LLFF dataset. We highlight the difference with colored patches. 5.3 Ablation Studies We conduct ablation experiments on the LLFF dataset. Quantitative results are shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.CV 2026-08 conditional novelty 6.0 of 10

    CLEAR performs single-stage joint training of a 3D Gaussian scene under low- and high-resolution supervision, using conflict-aware gradient correction and evidence-guided detail routing to reach state-of-the-art spars...

Reference graph

Works this paper leans on

69 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. 2021. Mip-nerf: A multiscale represen- tation for anti-aliasing neural radiance fields. InProceedings of the IEEE/CVF international conference on computer vision. 5855–5864

  2. [2]

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. 2022. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. 5470–5479

  3. [3]

    Mario Botsch, Alexander Hornung, Matthias Zwicker, and Leif Kobbelt. 2005. High-quality surface splatting on today’s GPUs. InProceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005.IEEE, 17–141

  4. [4]

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. 2022. TensoRF: Tensorial Radiance Fields. InEuropean Conference on Computer Vision (ECCV)

  5. [5]

    Rui Chen, Songfang Han, Jing Xu, and Hao Su. 2019. Point-based multi-view stereo network. InProceedings of the IEEE/CVF international conference on com- puter vision. 1538–1547

  6. [6]

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. 2024. Hac: Hash-grid assisted context for 3d gaussian splatting compression. InEuro- pean Conference on Computer Vision. Springer, 422–438

  7. [7]

    Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. 2024. Depth-regularized optimization for 3d gaussian splatting in few-shot images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 811–820

  8. [8]

    Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. 2022. Fov-nerf: Foveated neural radi- ance fields for virtual reality.IEEE Transactions on Visualization and Computer Graphics28, 11 (2022), 3854–3864

Show all 69 references
  1. [9]

    Xiaobiao Du, Yida Wang, and Xin Yu. 2024. Mvgs: Multi-view-regulated gaussian splatting for novel view synthesis.arXiv preprint arXiv:2410.02103(2024)

  2. [10]

    Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Bao- quan Chen. 2024. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes.arXiv preprint arXiv:2402.03307(2024)

  3. [11]

    Zhiwen Fan, Kairun Wen, Wenyan Cong, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. 2024. InstantSplat: Sparse-view Gaussian Splatting in Seconds. arXiv:2403.20309 [cs.CV]

  4. [12]

    Xiang Feng, Yongbo He, Yubo Wang, Chengkai Wang, Zhenzhong Kuang, Jiajun Ding, Feiwei Qin, Jun Yu, and Jianping Fan. 2024. ZS-SRT: An efficient zero-shot super-resolution training method for Neural Radiance Fields.Neurocomputing 590 (2024), 127714

  5. [13]

    Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Zhenzhong Kuang, Yu Jun, Jianping Fan, et al. 2024. SRGS: Super-Resolution 3D Gaussian Splatting.arXiv preprint arXiv:2404.10318(2024)

  6. [14]

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. 2022. Plenoxels: Radiance fields without neural net- works. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5501–5510

  7. [15]

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. 2018. Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG)37, 6 (2018), 1–15

  8. [16]

    Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. 2021. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision. 5875– 5884

  9. [17]

    Jintong Hu, Bin Xia, Bin Chen, Wenming Yang, and Lei Zhang. 2024. GaussianSR: High Fidelity 2D Gaussian Splatting for Arbitrary-Scale Image Super-Resolution. arXiv preprint arXiv:2407.18046(2024)

  10. [18]

    Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. 2023. Refsr-nerf: Towards high fidelity and super resolution view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8244–8253

  11. [19]

    Yiming Huang, Beilei Cui, Long Bai, Ziqi Guo, Mengya Xu, and Hongliang Ren. 2024. Endo-4dgs: Distilling depth ranking for endoscopic monocular scene reconstruction with 4d gaussian splatting.arXiv preprint arXiv:2401.16416(2024)

  12. [20]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  13. [21]

    Diederik P Kingma. 2013. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114(2013)

  14. [22]

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. 2017. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG)36, 4 (2017), 1–13

  15. [23]

    Hyun-kyu Ko, Dongheok Park, Youngin Park, Byeonghyeon Lee, Juhee Han, and Eunbyung Park. 2024. Sequence Matters: Harnessing Video Models in Super-Resolution.arXiv preprint arXiv:2412.11525(2024)

  16. [24]

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park

  17. [25]

    Jie Long Lee, Chen Li, and Gim Hee Lee. 2024. DiSR-NeRF: Diffusion-Guided View-Consistent Super-Resolution NeRF. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition. 20561–20570

  18. [26]

    Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu

  19. [27]

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. 2021. Swinir: Image restoration using swin transformer. InProceedings of the IEEE/CVF international conference on computer vision. 1833–1844

  20. [28]

    Chien-Yu Lin, Qichen Fu, Thomas Merth, Karren Yang, and Anurag Ranjan

  21. [29]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20775–20785

  22. [30]

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt

  23. [31]

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. 2024. Scaffold-gs: Structured 3d gaussians for view-adaptive render- ing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20654–20664

  24. [32]

    InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Fastsr-nerf: Improving nerf efficiency on consumer devices with a simple super-resolution pipeline. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 6036–6045

  25. [33]

    Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. 2022. Efficient neural radiance fields for interactive free-viewpoint video. InSIGGRAPH Asia 2022 Conference Papers. 1–9

  26. [34]

    Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. 2022. In- stant neural graphics primitives with a multiresolution hash encoding.ACM transactions on graphics (TOG)41, 4 (2022), 1–15

  27. [35]

    Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. 2022. Extracting triangular 3d models, materials, and lighting from images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition....

  28. [36]

    Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. 2024. Com- pressed 3d gaussian splatting for accelerated novel view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10349– 10358. Shiyun Xie, Zhiru Wang, Yinghao Zhu, X...

  29. [37]

    Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalan- tari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. 2019. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines.ACM Transactions on Graphics (ToG)38, 4 (2019), 1–14

  30. [38]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis.Commun. ACM65, 1 (2021), 99–106

  31. [39]

    Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. 2024. Revising densifi- cation in gaussian splatting. InEuropean Conference on Computer Vision. Springer, 347–362

  32. [40]

    Johannes L Schonberger and Jan-Michael Frahm. 2016. Structure-from-motion revisited. InProceedings of the IEEE conference on computer vision and pattern recognition. 4104–4113

  33. [41]

    Jianxiong Shen, Antonio Agudo, Francesc Moreno-Noguer, and Adria Ruiz. 2022. Conditional-flow nerf: Accurate 3d modelling with reliable uncertainty quantifi- cation. InEuropean Conference on Computer Vision. Springer, 540–557

  34. [42]

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. 2022. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988(2022)

  35. [43]

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai

  36. [44]

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898(2024)

  37. [45]

    Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. 2024. Neurad: Neural rendering for au- tonomous driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14895–14904

  38. [46]

    Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. 2024. Dn-splatter: Depth and normal priors for gaussian splatting and meshing.arXiv preprint arXiv:2403.17822(2024)

  39. [47]

    Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi- Min Hu. 2022. Nerf-sr: High quality neural radiance fields using supersampling. InProceedings of the 30th ACM International Conference on Multimedia. 6445– 6454

  40. [48]

    Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. 2021. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d repre- sentations. In2021 International Conference on 3D Vision (3DV). IEEE, 972–981

  41. [49]

    Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Früstück. 2024. SuperGaussian: Repurposing Video Models for 3D Super Resolution.arXiv preprint arXiv:2406.00609(2024)

  42. [50]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional net- works for large-scale image recognition.arXiv preprint arXiv:1409.1556(2014)

  43. [51]

    2024.Sparsegs: Real-time 360°sparse view synthesis using gaussian splatting

    Haolin Xiong. 2024.Sparsegs: Real-time 360°sparse view synthesis using gaussian splatting. Master’s thesis. University of California, Los Angeles

  44. [52]

    Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. 2024. Multi-scale 3d gaussian splatting for anti-aliased rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20923–20931

  45. [53]

    Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. 2024. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting.arXiv preprint arXiv:2402.15870(2024)

  46. [54]

    Chan, and Chen Change Loy

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. 2024. Exploiting Diffusion Prior for Real-World Image Super- Resolution. (2024)

  47. [55]

    Zhiru Wang, Shiyun Xie, Chengwei Pan, and Guoping Wang. 2024. SpecGaussian with Latent Features: A High-quality Modeling of the View-dependent Appear- ance for 3D Gaussian Splatting. InProceedings of the 32nd ACM International Conference on Multimedia. 6270–6278

  48. [56]

    Zhengxue Wang, Zhiqiang Yan, and Jian Yang. 2024. Sgnet: Structure guided network via gradient-frequency awareness for depth map super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 5823–5831

  49. [57]

    Youngho Yoon and Kuk-Jin Yoon. 2023. Cross-guided optimization of radiance fields with multi-view image super-resolution for high-resolution novel view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12428–12438

  50. [58]

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip-splatting: Alias-free 3d gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 19447–19456

  51. [59]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  52. [60]

    Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. 2023. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting.arXiv preprint arXiv:2310.10642(2023)

  53. [61]

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. 2018. Mvsnet: Depth inference for unstructured multi-view stereo. InProceedings of the European conference on computer vision (ECCV). 767–783

  54. [62]

    Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine- Hornung. 2019. Differentiable surface splatting for point-based geometry pro- cessing.ACM Transactions on Graphics (TOG)38, 6 (2019), 1–14

  55. [67]

    Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. 2024. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In European Conference on Computer Vision. Springer, 326–342

  56. [68]

    Zhe Zhang, Rui Peng, Yuxi Hu, and Ronggang Wang. 2023. Geomvsnet: Learning multi-view stereo with geometry perception. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 21508–21518

  57. [69]

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. 2023. FSGS: Real-Time Few-Shot View Synthesis using Gaussian Splatting. arXiv:2312.00451 [cs.CV]

  58. [2018]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition. 586–595

  59. [2020]

    Neural sparse voxel fields.Advances in Neural Information Processing Systems33 (2020), 15651–15663

  60. [2023]

    Graph.42, 4 (2023), 139–1

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Trans. Graph.42, 4 (2023), 139–1

  61. [2024]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Compact 3d gaussian representation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21719–21728

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.