Pith. sign in

REVIEW 3 major objections 4 minor 71 references

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

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CLEAR proposes the first unified single-stage framework for sparse-view 3D Gaussian super-resolution, jointly optimizing authentic low-resolution observations and external high-resolution priors within one Gaussian field.

desk verdict CLEAR is a coherent single-stage method for sparse-view 3DGS super-resolution with genuine efficiency gains, but the central LR-anchor premise is asserted rather than isolated in experiments, and the SOTA claims rest on small single-run margins. read the letter →

arxiv 2608.02206 v1 pith:5XSLA3TL submitted 2026-08-03 cs.CV

Hantang Li , Qiang Zhu , Xiandong Meng , Debin Zhao , Xiaopeng Fan This is my paper
classification cs.CV
keywords sparse-view3DGaussiansplattingsuper-resolutionunifiedsingle-stageoptimizationgradientconflictconflict-awareevidence-guidedroutingdensificationstabilization
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

This paper aims to establish that sparse-view 3D Gaussian super-resolution can be done in a single optimization stage, without the two-stage LR-reconstruction-then-HR-refinement paradigm that prior methods use. Instead of transferring one Gaussian field into another, CLEAR keeps one Gaussian field and supervises it at both low and high resolution: the true LR input images anchor the geometry, while HR references from an external image super-resolution model supply detail. The central problem this creates is that LR and HR gradients often pull a Gaussian in opposite directions, and the paper shows this conflict worsens as views become sparser. Its proposed fix is a per-Gaussian conflict-aware optimizer that always keeps the LR gradient and only softly corrects the HR gradient when it strongly disagrees, plus an evidence-routing mechanism that decides which high-frequency cues are trustworthy. If correct, the paper's numbers suggest the single-stage design not only avoids error accumulation from stage transfer but also trains roughly forty percent faster on a standard benchmark, producing the best reports on three 4x super-resolution datasets.

What carries the argument

The load-bearing object is the per-Gaussian gradient pair (LR and HR) and its cosine similarity, used as a conflict detector. The policy is asymmetric: the LR gradient is always fully kept, while the HR gradient is corrected only when its cosine with the LR gradient falls below a negative tolerance (the paper uses a tolerance threshold of 0.05), with correction strength modulated by an 'evidence' weight that combines patch-level reliability (how well the HR reference stays consistent with the LR image after down/up-sampling) and detail demand. The same evidence scores also weight the high-frequency Haar gradient and drive Gaussian densification, so that trustworthy details are injected and u

What would settle it

Train CLEAR on a scene while artificially corrupting the LR supervision (e.g., strong gaussian noise or occlusion) while keeping HR references clean; if the anchored gradients still produce geometrically faithful novel views, then LR-anchor reliability is not the active mechanism. Alternatively, run a two-stage baseline initialized from ground-truth LR geometry; if it matches or beats CLEAR, then the claim that stage-wise transfer is the main error source collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that gradient conflict between authentic LR supervision and external HR priors is measurable per Gaussian and can be reconciled explicitly. For each Gaussian, the method computes the cosine similarity between the LR gradient and the HR gradient; when the cosine falls below a negative tolerance, the HR gradient is softly corrected by partially removing its component against the LR gradient, with the correction strength modulated by patch evidence of reliability and detail demand. The same evidence scores are lifted from patches into Gaussian space and used to weight the high-frequency detail gradient and to guide densification. The result is a single

Load-bearing premise

The whole conflict-correction scheme assumes the authentic LR gradient is always a reliable guide for geometry, but under sparse views the LR optimization itself is geometrically ambiguous, and nothing in the paper validates that reliability or isolates what happens when it fails.

Editorial extensions

If this is right

  • If CLEAR is correct, the two-stage Gaussian-transfer paradigm is unnecessary; future sparse-view super-resolution methods can be designed as single-field optimizations, avoiding a class of error-accumulation bugs.
  • The conflict ratio (proportion of Gaussians with opposing LR/HR gradients) provides a training-time diagnostic; monitoring it may reveal when external priors and real observations diverge.
  • The reported roughly 40% shorter training time on LLFF indicates that removing stage-wise transfer yields a concrete efficiency gain, not just a quality gain.
  • Because the same Gaussian geometry is shared across scales, geometric fidelity learned from LR views directly constrains HR appearance, suggesting the unified design is well suited to very sparse inputs.
  • Best-reported PSNR/SSIM/LPIPS on three 4x benchmarks (LLFF 3-view, Blender 8-view, Mip-NeRF360 24-view) suggest the approach generalizes across synthetic and real scenes.

Reading between the lines

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

  • The explicit assumption that LR gradients are reliable anchors is untested; a direct comparison against a geometry-only or depth-prior baseline would show whether the anchor is necessary or whether the HR prior alone could support geometry.
  • The evidence-routing scores are soft per-patch attention values; they could be reused as confidence maps for downstream tasks such as view-consistent editing, uncertainty estimation, or automatic selection of which views to request for densification.
  • The same conflict-aware gradient merge could be applied to other multi-loss radiance-field problems (e.g., depth-prior plus photometric losses, or semantic plus appearance losses), where per-primitive gradient disagreement is similarly measurable.
  • The detached mid-training anchor is an early-bird style stabilization; an adaptive anchor schedule tied to the conflict-ratio plateau could make the method more robust to scene difficulty.
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

3 major / 4 minor

Summary. The paper proposes CLEAR, a unified single-stage 3D Gaussian Splatting framework for sparse-view 4x super-resolution. In contrast to prior two-stage pipelines (e.g., S2Gaussian), CLEAR jointly optimizes authentic low-resolution (LR) photometric loss and high-resolution (HR) supervision from a frozen image SR model within one Gaussian field. To reconcile the two supervision signals, it introduces Gaussian-wise Conflict-Aware Optimization (Eqs. 6-8), which treats the LR gradient as a reliable anchor and applies an evidence-conditioned soft correction to conflicting HR gradients. An evidence-guided Patch-to-Gaussian routing mechanism (Eqs. 9-14) lifts patch-level reliability and detail demand into Gaussian space to route HF gradients and guide densification. A sparse-view stabilization component (Eqs. 15-16) uses shared Gaussian dropout and detached mid-training anchoring. Experiments on LLFF (3 views), Blender (8 views), and Mip-NeRF 360 (24 views) report consistent improvements in PSNR/SSIM/LPIPS over existing methods, and the training-time comparison shows about a 40% reduction over S2Gaussian.

Significance. If the central claims hold, CLEAR is a meaningful step: a single-stage formulation for sparse-view 3D Gaussian super-resolution that avoids stage-wise error accumulation and is substantially more efficient than the two-stage baseline. The progressive ablation (Table 5) is a strength: each component improves all three metrics while reducing the measured LR-HR gradient conflict ratio, and the conflict-ratio diagnostic in Fig. 5 is a useful monitoring tool. The paper also gives a fairly complete description of the architecture and training details. However, the two load-bearing claims — that LR gradients are reliable anchors under very sparse views, and that CLEAR consistently achieves state-of-the-art quality — are not yet sufficiently supported. The SOTA comparisons are single-run point estimates with no variance reporting, and the LR-anchor premise receives no dedicated validation. The paper is therefore scientifically promising but currently falls short of its strongest claims.

major comments (3)
  1. [Gaussian-Wise Conflict-Aware Optimization, Eqs. (6)-(8); Fig. 1(b); Table 1] The core mechanism assumes that the LR gradient is a reliable anchor: Eq. (8) keeps g^LR unconditionally, and Eq. (7) only modifies the HR gradient. Under the test conditions this assumption is not self-evident. Plain 3DGS collapses to 12.51 dB PSNR on LLFF with 3 views (Table 1), and the paper's own Fig. 1(b) shows that gradient conflicts become more severe as views decrease. The LR-only optimization is therefore geometrically ambiguous, and a wrong LR gradient can suppress exactly the HR corrections that would fix geometry. The ablation in Table 5 only compares against naive joint training; it does not test symmetric correction, LR downweighting, or the addition of a geometry prior such as depth regularization. Moreover, the paper claims 'superior geometric fidelity' but reports no geometry metric. I request an experiment that isolates LR-gradient reliability — e.g., comparing Eq. (7)-
  2. [Tables 1-3 and Table 5] All reported quantitative results are single-run point estimates. There are no error bars, no multiple seeds, and no significance tests. The PSNR advantages over the reproduced S2Gaussian are +0.25 dB on LLFF, +0.48 dB on Blender, and +0.23 dB on Mip-NeRF 360 — margins that can easily fall within optimization noise for 3DGS-based methods. The ablation gains in Table 5 are likewise single-run. To support the 'consistently achieves state-of-the-art' claim, the authors should report mean and standard deviation over at least three seeds (or a per-scene breakdown) and state whether the improvement is consistent across scenes. This is especially important because the second-best results are only 0.2-0.5 dB behind in PSNR.
  3. [Eq. (7)] The sign convention in the soft-correction operator is confusing and appears to contradict the text. When the indicator I[c_i < -τ_c] is active, d_i is negative by construction (since c_i < -τ_c implies <g^LR,g^HR> + τ_c ||g^LR|| ||g^HR|| < 0). The update e^HR = g^HR - ω I d_i / ||g^LR||^2 g^LR therefore subtracts a negative vector, i.e., it adds an LR-aligned component rather than 'removing' the opposing HR component. The description says the operator 'removes only destructive HR components', which does not match the algebra as written. Please clarify whether the intended operation is a soft rotation toward g^LR, a projection with a compatibility margin, or a different correction, and adjust the equation or the text accordingly.
minor comments (4)
  1. [Eq. (16)] The stabilizing loss L_stab is described as 'attribute-weighted' via W, but W is never defined. Please specify how W weights each attribute (position, scale, rotation, appearance, opacity).
  2. [Implementation Details] The λ_HR schedule is only partially specified: it has a linear warm-up from iteration 500 to 3K and 'decay after 5K', but the actual value or decay curve in the 3K-5K interval and after 5K is not given. Please provide the full schedule.
  3. [Table 4] Training time is reported as an average over all scenes with no spread. Please include per-scene values or standard deviation, and state whether both methods were run under identical hardware and software conditions.
  4. [Ablation section] Minor typo: 'a effective balance' should be 'an effective balance'. Also, the paragraph on P2G states 'achieves a 0.37dB PSNR gain' from the baseline; in Table 5 it is +0.37 dB over the previous row (GCO), not from the baseline; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central benchmarks and ablations are external, self-citations are contextual, and no prediction reduces to a fitted parameter.

full rationale

No significant circularity is present. The central claim—that CLEAR outperforms prior sparse-view 3D Gaussian super-resolution methods on LLFF/Blender/Mip-NeRF360 4x benchmarks—is established by external quantitative comparisons (Tabs. 1–3) and ablations (Tab. 5) against public baselines and a reproduced S2Gaussian. The proposed mechanism (Eqs. 6–8, 9–14, 15–16) is an architectural/training design; no parameter is fitted to a target quantity and then reused as a prediction. The authors' self-citations (Li et al. 2026b,a) appear only in related-work lists as examples of sparse-view reconstruction methods that 'fail to recover high-frequency details'; they do not supply a uniqueness theorem, an ansatz, or a load-bearing premise of CLEAR, so they do not constitute circular support. The internally reported 'Conflict' ratio is a diagnostic, not a benchmark prediction, and the ablation's PSNR/SSIM/LPIPS gains are independent of that metric. The assumption that LR gradients are 'reliable anchors' under sparse views is unvalidated and is a genuine correctness risk, but an unverified premise is not circularity.

Assumptions & free parameters 9 free parameters · 5 assumptions · 1 invented entities

The framework introduces no speculative physical entities; its load-bearing weight is carried by hand-chosen hyperparameters (loss weights, thresholds, routing bounds) and three domain beliefs: ResShift HR references are trustworthy when consistent with the LR input, LR gradients are reliable geometry anchors under sparse views, and the soft-correction form (Eq. 7) removes only destructive HR components. These beliefs are supported only by ablation gains and the conflict-ratio diagnostic, with no external validation or formal analysis.

free parameters (9)
  • lambda_HR = 1.0 (linear warm-up iteration 500 to 3K, decay after 5K)
    Balances LR vs HR supervision in Eq. (8); chosen by hand per benchmark.
  • tau_c = 0.05 (correction scale factor bounded [0.05, 0.75])
    Conflict tolerance threshold in Eqs. (6)-(7); hand-set, controls how aggressively opposing HR gradients are softened.
  • tau_r = 0.05
    Reliability sensitivity in Eq. (9); controls how quickly SR patches are distrusted for routing.
  • q_min = 0.25
    Minimum routing weight in Eq. (12); preserves a floor of high-frequency supervision everywhere.
  • lambda_HF = 0.2
    Weight of the Haar high-frequency loss L_HF in Eq. (13).
  • lambda_stab = 0.01
    Weight of the detached anchor loss L_stab in Eq. (16).
  • maximum Gaussian dropout rate = 0.2
    Shared dropout level in Eq. (15), for sparse-view stabilization.
  • densification schedule and anchor iteration t_a = densify every 100 iterations from 500 to 5K; t_a after structure stabilizes
    SVS schedule; the criterion for 'structure stabilizes' is not formally defined.
  • lambda_ssim and lambda_HR_ssim = unspecified in text
    Photometric/SSIM balance weights in Eqs. (4)-(5); values not stated.
assumptions (5)
  • domain assumption Frozen ResShift HR references are trustworthy wherever D(I_HR) approximately equals I_LR (downsampling-consistency score).
    P2G routing (Eqs. 9-13) gates all HR detail on this evidence score; no multi-view consistency check of the 2D SR outputs is performed.
  • domain assumption LR gradients are reliable anchors for geometry under sparse views.
    Eq. (8) keeps g_i_LR unconditionally and only softens opposing HR gradients. The paper's own Fig. 1(b) shows conflicts rising with sparsity but never measures whether the LR gradient direction itself is correct.
  • ad hoc to paper The soft-correction operator (Eq. 7) removes only destructive HR components and preserves beneficial cross-resolution corrections.
    Heuristic margin-projection form; no derivation, no convergence or bias analysis is provided.
  • domain assumption Sharing the dropout mask z_t between LR and HR renderings preserves consistency rather than injecting artifacts (Eq. 15).
    Borrowed from DropGaussian/DropoutGS; assumed to regularize co-adaptation without breaking LR-HR alignment.
  • standard math 3DGS differentiable rasterization, adaptive density control, and Mip-Splatting scale-aware rasterization carry over unchanged.
    The method builds on Kerbl et al. 2023 and Yu et al. 2024b primitives; correctness of those pipelines is taken as given.
invented entities (1)
  • View-independent HR detail residual f_det
    purpose: Adds per-Gaussian high-frequency appearance capacity at HR scale only (Eq. 3); updated solely by routed HR detail gradients (Eq. 13).
    An internal learnable parameter, not a physically testable entity; its effect is observable only inside this model's rendered outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLEAR: Conflict-aware Learning via Evidence-guided Adaptive Routing for Unified Sparse-View 3D Gaussian Super-Resolution." pith.science (2026). https://pith.science/paper/5XSLA3TL

@misc{pith2026260802206,
  author       = {Pith},
  title        = {Pith review of: CLEAR: Conflict-aware Learning via Evidence-guided Adaptive Routing for Unified Sparse-View 3D Gaussian Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5XSLA3TL}},
  note         = {Machine review of arXiv:2608.02206}
}
abstract

Sparse-view 3D Gaussian Splatting Super-resolution is highly challenging since the sparse and low-resolution (LR) inputs lack sufficient geometric and high-frequency information for accurate reconstruction. To achieve high-quality reconstruction, existing sparse-view super-resolution methods adhere to two-stage pipeline that performs LR Gaussian reconstruction and then high-resolution (HR) Gaussian refinement, which directly results in stage-wise Gaussian transfer and reconstruction error accumulation. To this end, we propose CLEAR, a Conflict-aware Learning via Evidence-guided Adaptive Routing, as the first unified single-stage framework for Sparse-view 3D Gaussian Splatting Super-resolution. Specifically, CLEAR performs joint the optimization of authentic LR observations and external HR priors within a unified Gaussian representation. To mitigate the gradient conflicts introduced by sparse supervision during training, we propose a Gaussian-wise conflict-aware optimization strategy that regards the LR gradient as a reliable anchor and applies evidence-conditioned soft correction only to severe HR conflicts. Moreover, to recover high-frequency details, we introduce an evidence-guided Patch-to-Gaussian routing mechanism which estimates patch reliability and detail demand, lifts them into Gaussian space, and selectively routes high-frequency gradients and densification. Finally, we employ shared Gaussian dropout and a detached mid-training anchoring to enhance the robustness of training framework. Extensive experiments on both synthetic and real-world $4\times$ super-resolution benchmarks demonstrate that CLEAR consistently achieves state-of-the-art rendering quality and superior geometric fidelity.

Figures

Figures reproduced from arXiv: 2608.02206 by the authors.

Figure 1
Figure 1. Motivation and Performance. (a) Unlike existing two-stage methods, CLEAR is the first unified single-stage framework that jointly optimizes LR and HR supervision within a single Gaussian field. (b) Sparse-view observations intensify Gaussian-wise LR and HR gradient conflicts, leading to severe novel-view artifacts. (c) The radar plot summarizes PSNR and LPIPS across LLFF, Blender, and Mip-NeRF 360, where CLEAR achie… view at source ↗
Figure 2
Figure 2. Overview of the proposed CLEAR framework. Given sparse LR observations, CLEAR optimizes a Gaussian field along a unified single-stage optimization, where (a) Gaussian-wise Conflict-aware Optimization reconciles LR and HR gradient conflicts, (b) Evidence-guided Patch-to-Gaussian Routing transfers reliable detail cues from SR patches to visible Gaussians, and (c) Sparse-view Stabilization regularizes the Gaussian fiel… view at source ↗
Figure 3
Figure 3. Qualitative comparison on LLFF 4× super-resolution. Zoom in for the highlighted regions. Sparse-View Stabilization Sparse-view optimization is prone to Gaussian co-adaptation and late-stage geometric drift. To mitigate this, we first apply the same dropout mask to both rendering scales: z t i = Bernoulli(1 − pt) 1 − pt , Ibv r = Rr [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on Blender and Mip-NeRF 360 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Incremental analysis of CLEAR. Top: Gaussian￾wise LR and HR gradient conflict ratio averaged over eight LLFF scenes during training. Bottom: Novel-view results of different ablation models. Our components progressively reduce gradient conflicts and improve reconstructi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 10 linked inside Pith

  1. [1]

    and Mildenhall, Ben and Tancik, Matthew and Hedman, Peter and Martin-Brualla, Ricardo and Srinivasan, Pratul P

    Barron, Jonathan T. and Mildenhall, Ben and Tancik, Matthew and Hedman, Peter and Martin-Brualla, Ricardo and Srinivasan, Pratul P. ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  2. [2]

    and Mildenhall, Ben and Verbin, Dor and Srinivasan, Pratul P

    Barron, Jonathan T. and Mildenhall, Ben and Verbin, Dor and Srinivasan, Pratul P. and Hedman, Peter ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  3. [3]

    European Conference on Computer Vision (ECCV) ,pages=

    Chen, Anpei and Xu, Zexiang and Geiger, Andreas and Yu, Jingyi and Su, Hao ,title=. European Conference on Computer Vision (ECCV) ,pages=

  4. [4]

    arXiv preprint arXiv:2401.03890 ,year=

    Chen, Guikun and Wang, Wenguan ,title=. arXiv preprint arXiv:2401.03890 ,year=

    arXiv

  5. [5]

    International Conference on Machine Learning (ICML) ,year=

    Cheng, Kai and Long, Xiaoxiao and Yang, Kaizhi and Yao, Yao and Yin, Wei and Ma, Yuexin and Wang, Wenping and Chen, Xuejin ,title=. International Conference on Machine Learning (ICML) ,year=

  6. [6]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Chung, Jaeyoung and Oh, Jeongtaek and Lee, Kyoung Mu ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  7. [7]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Deng, Kangle and Liu, Andrew and Zhu, Jun-Yan and Ramanan, Deva ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  8. [8]

    arXiv preprint arXiv:2404.10318 ,year=

    Feng, Xiang and He, Yongbo and Wang, Yubo and Yang, Yan and Kuang, Zhenzhong and Jun, Yu and Fan, Jianping ,title=. arXiv preprint arXiv:2404.10318 ,year=

    arXiv

Show all 71 references
  1. [9]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Fridovich-Keil, Sara and Yu, Alex and Tancik, Matthew and Chen, Qinhong and Recht, Benjamin and Kanazawa, Angjoo ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  2. [10]

    and Kowalski, Marek and Johnson, Matthew and Shotton, Jamie and Valentin, Julien ,title=

    Garbin, Stephan J. and Kowalski, Marek and Johnson, Matthew and Shotton, Jamie and Valentin, Julien ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  3. [11]

    arXiv preprint arXiv:2304.13518 ,year=

    Han, Yuqi and Yu, Tao and Yu, Xiaohang and Wang, Yuwang and Dai, Qionghai ,title=. arXiv preprint arXiv:2304.13518 ,year=

    arXiv

  4. [12]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

    Asthana, Pranav and Hanson, Alex and Tu, Allen and Goldstein, Tom and Zwicker, Matthias and Varshney, Amitabh ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

  5. [13]

    AAAI Conference on Artificial Intelligence (AAAI) ,pages=

    Feng, Xiang and Zhong, Tieshi and Chang, Shuo and Wang, Weiliu and Wang, Chengkai and Chen, Yifei and Hu, Tongyu and Wang, Yuhe and Kuang, Zhenzhong and Yin, Xuefei and Zhu, Yanming ,title=. AAAI Conference on Artificial Intelligence (AAAI) ,pages=

  6. [14]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  7. [15]

    Advances in Neural Information Processing Systems (NeurIPS) ,volume=

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp ,title=. Advances in Neural Information Processing Systems (NeurIPS) ,volume=

  8. [16]

    IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

    H. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  9. [17]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Huang, Xudong and Li, Wei and Hu, Jie and Chen, Hanting and Wang, Yunhe ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  10. [18]

    Electronics Letters ,volume=

    Huynh-Thu, Quan and Ghanbari, Mohammed ,title=. Electronics Letters ,volume=

  11. [19]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering ,journal=

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. 3D Gaussian Splatting for Real-Time Radiance Field Rendering ,journal=

  12. [20]

    European Conference on Computer Vision (ECCV) , pages=

    Nerf: Representing scenes as neural radiance fields for view synthesis , author=. European Conference on Computer Vision (ECCV) , pages=

  13. [21]

    IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

    Kerr, Justin and Kim, Chung Min and Goldberg, Ken and Kanazawa, Angjoo and Tancik, Matthew ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  14. [22]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Lee, Jie Long and Li, Chen and Lee, Gim Hee ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  15. [23]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Li, Jiahe and Zhang, Jiawei and Bai, Xiao and Zheng, Jin and Ning, Xin and Zhou, Jun and Gu, Lin ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  16. [24]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Lu, Tao and Yu, Mulin and Xu, Linning and Xiangli, Yuanbo and Wang, Limin and Lin, Dahua and Dai, Bo ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  17. [25]

    and Ortiz-Cayon, Rodrigo and Kalantari, Nima Khademi and Ramamoorthi, Ravi and Ng, Ren and Kar, Abhishek ,title=

    Mildenhall, Ben and Srinivasan, Pratul P. and Ortiz-Cayon, Rodrigo and Kalantari, Nima Khademi and Ramamoorthi, Ravi and Ng, Ren and Kar, Abhishek ,title=. ACM Transactions on Graphics ,volume=

  18. [26]

    and Tancik, Matthew and Barron, Jonathan T

    Mildenhall, Ben and Srinivasan, Pratul P. and Tancik, Matthew and Barron, Jonathan T. and Ramamoorthi, Ravi and Ng, Ren ,title=. Communications of the ACM ,volume=

  19. [27]

    Instant Neural Graphics Primitives with a Multiresolution Hash Encoding ,journal=

    M. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding ,journal=

  20. [28]

    and Mildenhall, Ben and Sajjadi, Mehdi S

    Niemeyer, Michael and Barron, Jonathan T. and Mildenhall, Ben and Sajjadi, Mehdi S. M. and Geiger, Andreas and Radwan, Noha ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  21. [29]

    and Mildenhall, Ben ,title=

    Poole, Ben and Jain, Ajay and Barron, Jonathan T. and Mildenhall, Ben ,title=. arXiv preprint arXiv:2209.14988 ,year=

    arXiv

  22. [30]

    Vision Transformers for Dense Prediction ,booktitle=

    Ranftl, Ren. Vision Transformers for Dense Prediction ,booktitle=

  23. [31]

    arXiv preprint arXiv:2303.07937 ,year=

    Seo, Junyoung and Jang, Wooseok and Kwak, Min-Seop and Kim, Hyeonsu and Ko, Jaehoon and Kim, Junho and Kim, Jin-Hwa and Lee, Jiyoung and Kim, Seungryong ,title=. arXiv preprint arXiv:2303.07937 ,year=

    arXiv

  24. [32]

    and Wang, Shenlong and Fr

    Shen, Yuan and Ceylan, Duygu and Guerrero, Paul and Xu, Zexiang and Mitra, Niloy J. and Wang, Shenlong and Fr. SuperGaussian: Repurposing Video Models for 3D Super Resolution ,journal=

  25. [33]

    arXiv preprint arXiv:2309.16653 ,year=

    Tang, Jiaxiang and Ren, Jiawei and Zhou, Hang and Liu, Ziwei and Zeng, Gang ,title=. arXiv preprint arXiv:2309.16653 ,year=

    arXiv

  26. [34]

    and Srinivasan, Pratul P

    Verbin, Dor and Hedman, Peter and Mildenhall, Ben and Zickler, Todd and Barron, Jonathan T. and Srinivasan, Pratul P. ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  27. [35]

    Proceedings of the ACM International Conference on Multimedia ,pages=

    Wang, Chen and Wu, Xian and Guo, Yuan-Chen and Zhang, Song-Hai and Tai, Yu-Wing and Hu, Shi-Min ,title=. Proceedings of the ACM International Conference on Multimedia ,pages=

  28. [36]

    IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

    Wang, Guangcong and Chen, Zhaoxi and Loy, Chen Change and Liu, Ziwei ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  29. [37]

    Wang, Jianyi and Yue, Zongsheng and Zhou, Shangchen and Chan, Kelvin C. K. and Loy, Chen Change ,title=. International Journal of Computer Vision ,pages=

  30. [38]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Wang, Shuzhe and Leroy, Vincent and Cabon, Yohann and Chidlovskii, Boris and Revaud, J. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  31. [39]

    and Sheikh, Hamid R

    Wang, Zhou and Bovik, Alan C. and Sheikh, Hamid R. and Simoncelli, Eero P. ,title=. IEEE Transactions on Image Processing ,volume=

  32. [40]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Wu, Guanjun and Yi, Taoran and Fang, Jiemin and Xie, Lingxi and Zhang, Xiaopeng and Wei, Wei and Liu, Wenyu and Tian, Qi and Wang, Xinggang ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  33. [41]

    arXiv preprint arXiv:2312.00206 ,year=

    Xiong, Haolin and Muttukuru, Sairisheek and Upadhyay, Rishi and Chari, Pradyumna and Kadambi, Achuta ,title=. arXiv preprint arXiv:2312.00206 ,year=

    arXiv

  34. [42]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Yan, Zhiwen and Low, Weng Fei and Chen, Yu and Lee, Gim Hee ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  35. [43]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Yang, Jiawei and Pavone, Marco and Wang, Yue ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  36. [44]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Yu, Zehao and Chen, Anpei and Huang, Binbin and Sattler, Torsten and Geiger, Andreas ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  37. [45]

    arXiv preprint arXiv:2406.10111 ,year=

    Yu, Xiqian and Zhu, Hanxin and He, Tianyu and Chen, Zhibo ,title=. arXiv preprint arXiv:2406.10111 ,year=

    arXiv

  38. [46]

    Advances in Neural Information Processing Systems (NeurIPS) ,volume=

    Yue, Zongsheng and Wang, Jianyi and Loy, Chen Change ,title=. Advances in Neural Information Processing Systems (NeurIPS) ,volume=

  39. [47]

    and Shechtman, Eli and Wang, Oliver ,title=

    Zhang, Richard and Isola, Phillip and Efros, Alexei A. and Shechtman, Eli and Wang, Oliver ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  40. [48]

    IEEE Transactions on Image Processing ,year=

    Zheng, Xiaoyun and Liao, Liwei and Jiao, Jianbo and Gao, Feng and Wang, Ronggang ,title=. IEEE Transactions on Image Processing ,year=

  41. [49]

    Zhi, Shuaifeng and Laidlow, Tristan and Leutenegger, Stefan and Davison, Andrew J. ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  42. [50]

    European Conference on Computer Vision (ECCV) ,pages=

    Zhu, Zehao and Fan, Zhiwen and Jiang, Yifan and Wang, Zhangyang ,title=. European Conference on Computer Vision (ECCV) ,pages=

  43. [51]

    European Conference on Computer Vision (ECCV) ,pages=

    Zhang, Jiawei and Li, Jiahe and Yu, Xiaohan and Huang, Lei and Gu, Lin and Zheng, Jin and Bai, Xiao ,title=. European Conference on Computer Vision (ECCV) ,pages=

  44. [52]

    ACM Transactions on Graphics ,volume=

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. ACM Transactions on Graphics ,volume=

  45. [53]

    and Wang, Shenlong and Fr

    Shen, Yuan and Ceylan, Duygu and Guerrero, Paul and Xu, Zexiang and Mitra, Niloy J. and Wang, Shenlong and Fr. European Conference on Computer Vision (ECCV) ,year=

  46. [54]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Yoon, Youngho and Yoon, Kuk-Jin ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  47. [55]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

    Park, Hyunwoo and Ryu, Gun and Kim, Wonjun ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

  48. [56]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

    Xu, Yexing and Wang, Longguang and Chen, Minglin and Ao, Sheng and Li, Li and Guo, Yulan ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

  49. [57]

    arXiv preprint arXiv:2604.06739 ,year=

    Li, Hantang and Zhu, Qiang and Meng, Xiandong and Zhao, Debin and Fan, Xiaopeng ,title=. arXiv preprint arXiv:2604.06739 ,year=

    arXiv

  50. [58]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

    Fang, Shuangkang and Shen, I-Chao and Zhang, Xuanyang and Wang, Zesheng and Wang, Yufeng and Ding, Wenrui and Yu, Gang and Igarashi, Takeo ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

  51. [59]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Wan, Yecong and Shao, Mingwen and Cheng, Yuanshuo and Zuo, Wangmeng ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  52. [60]

    and Mildenhall, Ben and Verbin, Dor and Srinivasan, Pratul P

    Barron, Jonathan T. and Mildenhall, Ben and Verbin, Dor and Srinivasan, Pratul P. and Hedman, Peter ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  53. [61]

    International Conference on Machine Learning (ICML) ,pages=

    Cheng, Kai and Long, Xiaoxiao and Yang, Kaizhi and Yao, Yao and Yin, Wei and Ma, Yuexin and Wang, Wenping and Chen, Xuejin ,title=. International Conference on Machine Learning (ICML) ,pages=

  54. [62]

    IEEE Transactions on Pattern Analysis and Machine Intelligence ,pages=

    Ren, Kerui and Jiang, Lihan and Lu, Tao and Yu, Mulin and Xu, Linning and Ni, Zhangkai and Dai, Bo ,title=. IEEE Transactions on Pattern Analysis and Machine Intelligence ,pages=

  55. [63]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

    Zheng, Yulong and Jiang, Zicheng and He, Shengfeng and Sun, Yandu and Dong, Junyu and Zhang, Huaidong and Du, Yong ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

  56. [64]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

    Jang, Youngkyoon and P. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,year=

  57. [65]

    IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

    Zhao, Chen and Wang, Xuan and Zhang, Tong and Javed, Saqib and Salzmann, Mathieu ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,pages=

  58. [66]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

    Kong, Hanyang and Yang, Xingyi and Wang, Xinchao ,title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,pages=

  59. [67]

    IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) ,year=

    Topalo. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) ,year=

  60. [68]

    and Huang, Jia-Bin ,title=

    Chen, Yi-Ting and Liao, Ting-Hsuan and Guo, Pengsheng and Schwing, Alexander G. and Huang, Jia-Bin ,title=. IEEE/CVF International Conference on Computer Vision (ICCV) ,year=

  61. [69]

    arXiv preprint arXiv:2505.18649 ,year=

    Xie, Shiyun and Wang, Zhiru and Zhu, Yinghao and Wang, Xu and Pan, Chengwei and Dong, Xiwang ,title=. arXiv preprint arXiv:2505.18649 ,year=

    arXiv

  62. [70]

    arXiv preprint arXiv:2605.12072 ,year=

    Li, Hantang and Zhu, Qiang and Meng, Xiandong and Wang, Xingtao and Zhao, Debin and Fan, Xiaopeng ,title=. arXiv preprint arXiv:2605.12072 ,year=

    arXiv

  63. [71]

    Advances in Neural Information Processing Systems (NeurIPS) ,volume=

    Chen, Kangjie and Zhong, Yingji and Li, Zhihao and Lin, Jiaqi and Chen, Youyu and Qin, Minghan and Wang, Haoqian ,title=. Advances in Neural Information Processing Systems (NeurIPS) ,volume=

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.