arxiv: 2604.17965 · v1 · submitted 2026-04-20 · 💻 cs.CV

Recognition: unknown

MU-GeNeRF: Multi-view Uncertainty-guided Generalizable Neural Radiance Fields for Distractor-aware Scene

Wenjie Mu , Zhan Li , Chuanzhou Su , Xuanyi Shen , Ziniu Liu , Fan Lu , Yujian Mo , Junqiao Zhao

show 3 more authors

Tiantian Feng Chen Ye Guang Chen

Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:22 UTC · model grok-4.3

classification 💻 cs.CV

keywords generalizable neural radiance fieldsdistractor handlinguncertainty estimationmulti-view consistencytransient objectsneural renderingscene reconstruction

0 comments

The pith

MU-GeNeRF decomposes uncertainty into source-view and target-view components so generalizable NeRFs can suppress transient distractors without per-scene optimization.

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

The paper establishes that generalizable neural radiance fields can maintain high-quality reconstruction even when transient objects break cross-view consistency by estimating two distinct uncertainties. Source-view uncertainty measures structural discrepancies among the sparse input images, while target-view uncertainty flags anomalies specific to the novel view being rendered. These signals are combined in a heteroscedastic loss that down-weights corrupted rays during training. A sympathetic reader would care because everyday photographs almost always contain moving people or vehicles, and the method removes the need to either curate perfectly static scenes or train a separate model for each new environment.

Core claim

The central claim is that decomposing distractor awareness into source-view uncertainty, which captures structural discrepancies across source views caused by viewpoint changes or dynamic factors, and target-view uncertainty, which detects observation anomalies in the target image induced by transient distractors, allows their combination through a heteroscedastic reconstruction loss to adaptively modulate supervision, enabling more robust distractor suppression and geometric modeling than prior generalizable NeRF approaches.

What carries the argument

The two complementary uncertainty components (source-view discrepancies across inputs and target-view anomalies in the novel view) integrated via the heteroscedastic reconstruction loss that adaptively weights supervision.

If this is right

The approach surpasses existing generalizable NeRF methods on scenes containing transient objects.
Reconstruction quality becomes comparable to scene-specific distractor-free NeRFs.
Geometric modeling improves because supervision is modulated rather than uniformly corrupted.
Cross-scene generalization remains possible while handling dynamic elements that break consistency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same uncertainty split could reduce reliance on manual cleaning of training datasets for neural rendering.
Similar multi-view uncertainty checks might transfer to other tasks that need robust 3D modeling from imperfect image sets.

Load-bearing premise

That the proposed source-view and target-view uncertainties can be computed reliably from reconstruction errors and viewpoint discrepancies without misclassifying consistent static structures as distractors.

What would settle it

A controlled test scene with known static geometry that appears inconsistent across views due only to viewpoint change, where the method degrades reconstruction quality below a baseline GeNeRF that uses no uncertainty weighting.

Figures

Figures reproduced from arXiv: 2604.17965 by Chen Ye, Chuanzhou Su, Fan Lu, Guang Chen, Junqiao Zhao, Tiantian Feng, Wenjie Mu, Xuanyi Shen, Yujian Mo, Zhan Li, Ziniu Liu.

**Figure 1.** Figure 1: Proposed Multi-view Uncertainty-guided distractoraware methods. Two complementary uncertainties are estimated to support stable training in dynamic scenes. Source-view Uncertainty captures structural inconsistencies across source views arising from occlusion or dynamic factors, while Target-view Uncertainty localizes the spatial distribution of distractors within the target view. Together, they synergis… view at source ↗

**Figure 2.** Figure 2: Overall architecture of our method. Given multi-view inputs, the model projects sampled points from target-view rays onto source images {I s n} N n=1 and feature maps {F s n} N n=1, sampling color {c s n} N n=1 and feature information {f s n} N n=1. Combined with spatial information xk and dk, these are fed into an feed-forward network Gθ to predict each point’s color ck and source-view uncertainty β s k. … view at source ↗

**Figure 3.** Figure 3: Visualization results under different uncertainty modeling strategies. (b) Modeling only Target-view Uncertainty β t tends to misjudge static regions as distractors. (c) Modeling only Source-view Uncertainty β s fails to localize distractors in the target view. (d) Jointly modeling both β s and β t and weighted fused into βts under the multi-view uncertainty framework. (e) and (f) are the separate visualiz… view at source ↗

**Figure 4.** Figure 4: Qualitative comparison on On-the-go after fine-tuning. MuRF† implicitly models inter-view consistency, making it susceptible to transient distractors and causing its rendered results to exhibit distractor artifacts. Method Setting Corner Patio Spot Patio-High PSNR↑SSIM↑LPIPS↓ PSNR↑ SSIM↑LPIPS↓ PSNR↑ SSIM↑LPIPS↓ PSNR↑ SSIM↑LPIPS↓ ReTR (NIPS 2023) No optimized per-scene 16.33 0.545 0.541 16.39 0.418 0.568 1… view at source ↗

**Figure 5.** Figure 5: Qualitative comparison on RobustNeRF after finetuning. Method Statue Android PSNR↑SSIM↑ PSNR↑SSIM↑ ReTR (NIPS 2023) 16.84 0.500 18.63 0.531 MuRF (CVPR 2024) 12.88 0.337 11.94 0.370 MU-GeNeRF (Ours) 18.28 0.577 19.87 0.583 ReTR ft (NIPS 2023) 18.77 0.585 20.29 0.601 MuRF ft (CVPR 2024) 13.10 0.351 12.34 0.389 MuRF† ft (CVPR 2024) 13.51 0.366 12.93 0.401 UP-NeRF (NIPS 2023) 18.10 0.629 20.45 0.664 NeRF on-t… view at source ↗

**Figure 6.** Figure 6: Qualitative comparison of view rendering results under various ablation components. non-continuous distractors across frames. We select four scenes from On-the-go and two from RobustNeRF for finetuning and reporting, while the remaining scenes are used for generalization training. Baselines We compare it with several representative methods. ReTR [20] and MuRF [51] are both GeNeRFbased methods designed f… view at source ↗

**Figure 7.** Figure 7: Rendering results under varying distractor injection levels. The numeral preceding “-Distractors” denotes the count of source views (out of eight) that include distractors. then fine-tune for 60K iters on per-scene. 5.2. Quantitative and Qualitative Evaluation Tab. 1 and Tab. 2 present quantitative comparisons with several baselines. Before or after per-scene fine-tuning, our method consistently outperform… view at source ↗

**Figure 8.** Figure 8: Visualization Comparison of Different Uncertainty Modeling Strategies. w/o β β s (only) β t (only) βts GT Image Corner w/o β β s (only) β t (only) βts GT Image Corner w/o β β s (only) β t (only) βts GT Image Corner w/o β β s (only) β t (only) βts GT Image Corner w/o β β s (only) β t (only) βts GT Image Corner w/o β β s (only) β t (only) βts GT Image Corner w/o β β s (only) β t (only) βts GT Image Corner w/… view at source ↗

**Figure 9.** Figure 9: Distractor Suppression Performance under Different Uncertainty Modeling Strategies. w/o β denotes that no uncertainty is modeled. method consistently produces renderings with consistent structure. While slightly lower than the distractor-free baseline, the results remain visually acceptable. This robustness stems from our proposed distractor-aware mthods, which suppresses distractor regions through uncert… view at source ↗

read the original abstract

Generalizable Neural Radiance Fields (GeNeRFs) enable high-quality scene reconstruction from sparse views and can generalize to unseen scenes. However, in real-world settings, transient distractors break cross-view structural consistency, corrupting supervision and degrading reconstruction quality. Existing distractor-free NeRF methods rely on per-scene optimization and estimate uncertainty from per-view reconstruction errors, which are not reliable for GeNeRFs and often misjudge inconsistent static structures as distractors. To this end, we propose MU-GeNeRF, a Multi-view Uncertainty-guided distractor-aware GeNeRF framework designed to alleviate GeNeRF's robust modeling challenges in the presence of transient distractions. We decompose distractor awareness into two complementary uncertainty components: Source-view Uncertainty, which captures structural discrepancies across source views caused by viewpoint changes or dynamic factors; and Target-view Uncertainty, which detects observation anomalies in the target image induced by transient distractors.These two uncertainties address distinct error sources and are combined through a heteroscedastic reconstruction loss, which guides the model to adaptively modulate supervision, enabling more robust distractor suppression and geometric modeling.Extensive experiments show that our method not only surpasses existing GeNeRFs but also achieves performance comparable to scene-specific distractor-free NeRFs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MU-GeNeRF adds a source-target uncertainty split and heteroscedastic loss to handle distractors in feed-forward GeNeRFs, but the abstract supplies no numbers or ablations to show it works.

read the letter

The punchline here is that MU-GeNeRF introduces a source-view and target-view uncertainty split to suppress transient distractors in generalizable NeRFs, using a heteroscedastic loss to adjust supervision. This is new in the context of feed-forward GeNeRFs, where per-scene uncertainty methods don't apply directly. The paper correctly identifies that reconstruction errors in this setting mix viewpoint changes, dynamics, and model generalization issues. Breaking it into structural discrepancies across sources and anomalies in the target is a reasonable decomposition. It does well at targeting a real usability issue for these models in scenes with moving objects, which limits their deployment in robotics or casual photography. The soft spots are significant though. The abstract asserts that experiments show it beats other GeNeRFs and matches scene-specific methods, yet no numbers, ablations, or details on uncertainty estimation appear. This makes it hard to evaluate the claims. The stress-test point holds weight: without per-scene fitting, high errors might reflect the shared model's limitations on a new scene rather than distractors alone. If the loss downweights those, it risks weakening the learning signal for static content. The paper would need specific tests to show the split avoids this. This work is for NeRF researchers focused on robust, real-world reconstruction from sparse views. A reader in that area could pick up the uncertainty modeling idea and see if it fits their setup. I recommend putting it through peer review. The problem is worth solving and the approach is clearly motivated, but the evidence needs to be presented and scrutinized.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes MU-GeNeRF, a generalizable NeRF framework that incorporates two uncertainty components—source-view uncertainty capturing structural discrepancies across input views and target-view uncertainty detecting anomalies in the rendered target—to handle transient distractors. These are combined in a heteroscedastic reconstruction loss to adaptively down-weight corrupted supervision. The central claim is that this enables the model to surpass prior GeNeRF methods while matching the quality of per-scene distractor-free NeRFs.

Significance. If the uncertainty estimates can be shown to isolate transient distractors without attenuating supervision on static geometry, the work would meaningfully extend generalizable NeRFs to real-world scenes containing moving objects or occluders. The decomposition into complementary source- and target-view terms is a reasonable response to the limitations of per-scene uncertainty methods when applied to feed-forward architectures.

major comments (3)

[Abstract] Abstract: the performance claims ('surpasses existing GeNeRFs' and 'achieves performance comparable to scene-specific distractor-free NeRFs') are presented without any quantitative metrics, PSNR/SSIM values, dataset names, or references to tables/figures. This absence prevents evaluation of the magnitude of improvement and is load-bearing for the central contribution.
[§3] §3 (Method): the source-view uncertainty (structural discrepancies) and target-view uncertainty (observation anomalies) are defined from reconstruction errors and viewpoint differences. Because the model is a feed-forward GeNeRF rather than per-scene optimized, these errors can equally arise from imperfect generalization of the shared weights to the current scene; the manuscript provides no analysis, ablation, or diagnostic to show that the heteroscedastic loss does not suppress valid static content under this regime.
[§4] §4 (Experiments): no ablation isolating the contribution of each uncertainty term, no comparison of uncertainty maps against ground-truth distractor masks, and no discussion of how reconstruction-error thresholds are chosen or whether they vary with scene complexity. These omissions leave the robustness claims unverified.

minor comments (2)

[Abstract] The abstract would be strengthened by a single sentence listing the datasets and the key quantitative gains (e.g., average PSNR improvement).
[§3] Notation for the two uncertainty maps and the combined heteroscedastic loss should be introduced with explicit equations early in §3 to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and have revised the manuscript to strengthen the presentation of results, provide additional analysis on uncertainty behavior, and include requested ablations and diagnostics.

read point-by-point responses

Referee: [Abstract] Abstract: the performance claims ('surpasses existing GeNeRFs' and 'achieves performance comparable to scene-specific distractor-free NeRFs') are presented without any quantitative metrics, PSNR/SSIM values, dataset names, or references to tables/figures. This absence prevents evaluation of the magnitude of improvement and is load-bearing for the central contribution.

Authors: We agree that the abstract should include concrete quantitative support for the central claims. In the revised manuscript we have updated the abstract to report specific PSNR and SSIM gains on the LLFF and Blender datasets (with references to Tables 1 and 2) while preserving the original length constraints. revision: yes
Referee: [§3] §3 (Method): the source-view uncertainty (structural discrepancies) and target-view uncertainty (observation anomalies) are defined from reconstruction errors and viewpoint differences. Because the model is a feed-forward GeNeRF rather than per-scene optimized, these errors can equally arise from imperfect generalization of the shared weights to the current scene; the manuscript provides no analysis, ablation, or diagnostic to show that the heteroscedastic loss does not suppress valid static content under this regime.

Authors: This is a valid concern. While the two uncertainty terms are designed to capture distinct error sources (cross-view structural inconsistency versus target-view anomalies), we acknowledge that the original submission lacked explicit verification that generalization errors are not misclassified as distractors. We have added a new subsection (3.4) with an analysis on static scenes without transients, showing that both uncertainty maps remain low on consistent geometry and that removing the heteroscedastic weighting does not improve (and slightly degrades) reconstruction quality. Corresponding quantitative results and visualizations are now included. revision: yes
Referee: [§4] §4 (Experiments): no ablation isolating the contribution of each uncertainty term, no comparison of uncertainty maps against ground-truth distractor masks, and no discussion of how reconstruction-error thresholds are chosen or whether they vary with scene complexity. These omissions leave the robustness claims unverified.

Authors: We agree these elements would strengthen the experimental validation. The revised experiments section now contains (i) an ablation isolating source-view versus target-view uncertainty contributions, (ii) qualitative side-by-side comparisons of predicted uncertainty maps against available ground-truth distractor masks on the synthetic and real distractor datasets, and (iii) a brief discussion clarifying that thresholds are not manually set but emerge from the learned heteroscedastic variance parameters, which adapt per scene without explicit dependence on scene complexity. revision: yes

Circularity Check

0 steps flagged

No circularity; uncertainty terms are independently computed from input discrepancies

full rationale

The MU-GeNeRF framework introduces source-view structural discrepancy and target-view observation anomaly as two new uncertainty components, each derived directly from measurable reconstruction errors and viewpoint differences in the input views. These feed into a heteroscedastic loss that modulates supervision weights. No equation reduces the claimed distractor suppression or performance gains to a parameter fitted from the target result itself. The method extends prior GeNeRF architectures with these explicit, data-driven terms rather than renaming or self-defining the output. No load-bearing self-citation chains or imported uniqueness theorems appear in the derivation; empirical comparisons to baselines are presented as external validation. This is the normal case of an incremental architectural extension whose central claims remain falsifiable against held-out scenes.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are explicitly introduced or quantified in the abstract; the approach relies on standard NeRF assumptions about radiance fields and photometric consistency.

pith-pipeline@v0.9.0 · 5561 in / 1089 out tokens · 49487 ms · 2026-05-10T05:22:43.254057+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

60 extracted references · 8 canonical work pages · 1 internal anchor

[1]

Light- ning nerf: Efficient hybrid scene representation for au- tonomous driving

Junyi Cao, Zhichao Li, Naiyan Wang, and Chao Ma. Light- ning nerf: Efficient hybrid scene representation for au- tonomous driving. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16803–16809. IEEE, 2024. 1

2024
[2]

Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. InProceedings of the IEEE/CVF international conference on computer vision, pages 14124–14133, 2021. 2

2021
[3]

Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation

Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 19436–19446, 2024. 2

2024
[4]

Hallucinated neural radiance fields in the wild

Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 12943–12952, 2022. 2

2022
[5]

Explicit correspondence matching for generalizable neural radiance fields.IEEE Transactions on Pattern Analysis and Machine Intelligence,

Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields.IEEE Transactions on Pattern Analysis and Machine Intelligence,
[6]

Stereo radiance fields (srf): Learning view syn- thesis for sparse views of novel scenes

Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view syn- thesis for sparse views of novel scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7911–7920, 2021. 2

2021
[7]

Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts

Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 3193–3204, 2023. 1

2023
[8]

Go-nerf: Generating objects in neural ra- diance fields for virtual reality content creation.IEEE Trans- actions on Visualization and Computer Graphics, 2025

Peng Dai, Feitong Tan, Xin Yu, Yifan Peng, Yinda Zhang, and Xiaojuan Qi. Go-nerf: Generating objects in neural ra- diance fields for virtual reality content creation.IEEE Trans- actions on Visualization and Computer Graphics, 2025. 1

2025
[9]

Neural radiance fields (nerfs): A review and some recent developments.arXiv preprint arXiv:2305.00375, 2023

Mohamed Debbagh. Neural radiance fields (nerfs): A review and some recent developments.arXiv preprint arXiv:2305.00375, 2023. 1

work page arXiv 2023
[10]

Fov-nerf: Foveated neural radiance fields for virtual reality.IEEE Transactions on Visualization and Computer Graphics, 28(11):3854–3864, 2022

Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality.IEEE Transactions on Visualization and Computer Graphics, 28(11):3854–3864, 2022. 1

2022
[11]

Learning to render novel views from wide-baseline stereo pairs

Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitz- mann. Learning to render novel views from wide-baseline stereo pairs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970– 4980, 2023. 2

2023
[12]

Nerf: Neural radiance field in 3d vision, a comprehensive review,

Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review.arXiv preprint arXiv:2210.00379,

work page arXiv
[13]

Pc-nerf: Parent-child neural radiance fields using sparse lidar frames in autonomous driv- ing environments.IEEE Transactions on Intelligent Vehicles,

Xiuzhong Hu, Guangming Xiong, Zheng Zang, Peng Jia, Yuxuan Han, and Junyi Ma. Pc-nerf: Parent-child neural radiance fields using sparse lidar frames in autonomous driv- ing environments.IEEE Transactions on Intelligent Vehicles,
[14]

Refinedfields: Radiance fields refinement for unconstrained scenes.arXiv preprint arXiv:2312.00639, 3, 2023

Karim Kassab, Antoine Schnepf, Jean-Yves Franceschi, Laurent Caraffa, Jeremie Mary, and Val ´erie Gouet-Brunet. Refinedfields: Radiance fields refinement for unconstrained scenes.arXiv preprint arXiv:2312.00639, 2023. 2

work page arXiv 2023
[15]

Up- nerf: Unconstrained pose prior-free neural radiance field

Injae Kim, Minhyuk Choi, and Hyunwoo J Kim. Up- nerf: Unconstrained pose prior-free neural radiance field. Advances in Neural Information Processing Systems, 36: 68184–68196, 2023. 2, 6

2023
[16]

Semantic-aware occlusion filtering neural radiance fields in the wild.arXiv preprint arXiv:2303.03966, 2023

Jaewon Lee, Injae Kim, Hwan Heo, and Hyunwoo J Kim. Semantic-aware occlusion filtering neural radiance fields in the wild.arXiv preprint arXiv:2303.03966, 2023. 2

work page arXiv 2023
[17]

Improved neural radi- ance fields using pseudo-depth and fusion.arXiv preprint arXiv:2308.03772, 2023

Jingliang Li, Qiang Zhou, Chaohui Yu, Zhengda Lu, Jun Xiao, Zhibin Wang, and Fan Wang. Improved neural radi- ance fields using pseudo-depth and fusion.arXiv preprint arXiv:2308.03772, 2023. 1

work page arXiv 2023
[18]

Nerf-ms: Neural radiance fields with multi-sequence

Peihao Li, Shaohui Wang, Chen Yang, Bingbing Liu, We- ichao Qiu, and Haoqian Wang. Nerf-ms: Neural radiance fields with multi-sequence. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 18591– 18600, 2023. 2

2023
[19]

Dynibar: Neural dynamic image-based rendering

Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273– 4284, 2023. 2

2023
[20]

Retr: Modeling rendering via transformer for generalizable neural surface re- construction.Advances in Neural Information Processing Systems, 36:62332–62351, 2023

Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface re- construction.Advances in Neural Information Processing Systems, 36:62332–62351, 2023. 4, 6

2023
[21]

Uncertainty-aware 3d object-level mapping with deep shape priors

Ziwei Liao, Jun Yang, Jingxing Qian, Angela P Schoellig, and Steven L Waslander. Uncertainty-aware 3d object-level mapping with deep shape priors. In2024 IEEE international conference on robotics and automation (ICRA), pages 4082–
[22]

Efficient neural radiance fields for interactive free-viewpoint video

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

2022
[23]

Robust dynamic radiance fields

Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Jo- hannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 13–23, 2023. 2

2023
[24]

Nerf in the wild: Neural radiance fields for uncon- strained photo collections

Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duck- worth. Nerf in the wild: Neural radiance fields for uncon- strained photo collections. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210–7219, 2021. 1

2021
[25]

Ctnerf: Cross-time 9 transformer for dynamic neural radiance field from monocu- lar video.Pattern Recognition, 156:110729, 2024

Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time 9 transformer for dynamic neural radiance field from monocu- lar video.Pattern Recognition, 156:110729, 2024. 2

2024
[26]

Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 1

2021
[27]

Entangled view-epipolar information aggregation for generalizable neural radiance fields

Zhiyuan Min, Yawei Luo, Wei Yang, Yuesong Wang, and Yi Yang. Entangled view-epipolar information aggregation for generalizable neural radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4906–4916, 2024. 2

2024
[28]

High- fidelity 3d reconstruction via unified nerf-mesh optimization with geometric and color consistency.Pattern Recognition, page 112071, 2025

Rama Bastola Neupane, Kan Li, and Zhuqing Mao. High- fidelity 3d reconstruction via unified nerf-mesh optimization with geometric and color consistency.Pattern Recognition, page 112071, 2025. 1

2025
[29]

Cascaded and generalizable neural radiance fields for fast view synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):2758–2769, 2023

Phong Nguyen-Ha, Lam Huynh, Esa Rahtu, Jiri Matas, and Janne Heikkil¨a. Cascaded and generalizable neural radiance fields for fast view synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):2758–2769, 2023. 2

2023
[30]

Colnerf: collaboration for general- izable sparse input neural radiance field

Zhangkai Ni, Peiqi Yang, Wenhan Yang, Hanli Wang, Lin Ma, and Sam Kwong. Colnerf: collaboration for general- izable sparse input neural radiance field. InProceedings of the AAAI Conference on Artificial Intelligence, pages 4325– 4333, 2024. 2

2024
[31]

DINOv2: Learning Robust Visual Features without Supervision

Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4

work page internal anchor Pith review Pith/arXiv arXiv 2023
[32]

Ac- tivenerf: Learning where to see with uncertainty estimation

Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. Ac- tivenerf: Learning where to see with uncertainty estimation. InEuropean Conference on Computer Vision, pages 230–
[33]

Urban radiance fields

Konstantinos Rematas, Andrew Liu, Pratul P Srini- vasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12932–12942, 2022. 2

2022
[34]

Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild

Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931–8940, 2024. 1, 2, 6

2024
[35]

V olrecon: V olume rendering of signed ray distance functions for generalizable multi-view recon- struction

Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine S¨usstrunk. V olrecon: V olume rendering of signed ray distance functions for generalizable multi-view recon- struction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685– 16695, 2023. 4

2023
[36]

Robustnerf: Ig- noring distractors with robust losses

Sara Sabour, Suhani V ora, Daniel Duckworth, Ivan Krasin, David J Fleet, and Andrea Tagliasacchi. Robustnerf: Ig- noring distractors with robust losses. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20626–20636, 2023. 6

2023
[37]

Dynamon: Motion-aware fast and robust camera localization for dy- namic neural radiance fields.IEEE Robotics and Automation Letters, 2024

Nicolas Schischka, Hannah Schieber, Mert Asim Karaoglu, Melih Gorgulu, Florian Gr ¨otzner, Alexander Ladikos, Nas- sir Navab, Daniel Roth, and Benjamin Busam. Dynamon: Motion-aware fast and robust camera localization for dy- namic neural radiance fields.IEEE Robotics and Automation Letters, 2024. 2

2024
[38]

Boostmvsnerfs: Boost- ing mvs-based nerfs to generalizable view synthesis in large- scale scenes

Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai, Jie-Ying Lee, Chin-Yang Lin, and Yu-Lun Liu. Boostmvsnerfs: Boost- ing mvs-based nerfs to generalizable view synthesis in large- scale scenes. InACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 1

2024
[39]

Dyblurf: Dynamic neural radiance fields from blurry monocular video

Huiqiang Sun, Xingyi Li, Liao Shen, Xinyi Ye, Ke Xian, and Zhiguo Cao. Dyblurf: Dynamic neural radiance fields from blurry monocular video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7517–7527, 2024. 2

2024
[40]

Neural 3d reconstruction in the wild

Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3d reconstruction in the wild. InACM SIGGRAPH 2022 conference proceedings, pages 1–9, 2022. 2

2022
[41]

Global latent neu- ral rendering

Thomas Tanay and Matteo Maggioni. Global latent neu- ral rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19723– 19733, 2024. 2

2024
[42]

Block-nerf: Scalable large scene neural view synthesis

Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Prad- han, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 8248–8258, 2022. 2

2022
[43]

Neurad: Neural rendering for autonomous driving

Adam Tonderski, Carl Lindstr ¨om, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895–14904, 2024. 1

2024
[44]

Mega-nerf: Scalable construction of large- scale nerfs for virtual fly-throughs

Haithem Turki, Deva Ramanan, and Mahadev Satya- narayanan. Mega-nerf: Scalable construction of large- scale nerfs for virtual fly-throughs. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12922–12931, 2022. 2

2022
[45]

Is attention all that nerf needs?arXiv preprint arXiv:2207.13298, 2022

Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venu- gopalan, Zhangyang Wang, et al. Is attention all that nerf needs?arXiv preprint arXiv:2207.13298, 2022. 2

work page arXiv 2022
[46]

Ibr- net: Learning multi-view image-based rendering

Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibr- net: Learning multi-view image-based rendering. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690–4699, 2021. 2

2021
[47]

Ie-nerf: Exploring transient mask inpainting to enhance neural radiance fields in the wild.Neurocomputing, 618:129112, 2025

Shuaixian Wang, Haoran Xu, Yaokun Li, Jiwei Chen, and Guang Tan. Ie-nerf: Exploring transient mask inpainting to enhance neural radiance fields in the wild.Neurocomputing, 618:129112, 2025. 2

2025
[48]

Mp-nerf: More refined deblurred neu- ral radiance field for 3d reconstruction of blurred images

Xiaohui Wang, Zhenyu Yin, Feiqing Zhang, Dan Feng, and Zisong Wang. Mp-nerf: More refined deblurred neu- ral radiance field for 3d reconstruction of blurred images. Knowledge-Based Systems, 290:111571, 2024. 1

2024
[49]

Golf-nrt: Integrating global context and local geometry 10 for few-shot view synthesis

You Wang, Li Fang, Hao Zhu, Fei Hu, Long Ye, and Zhan Ma. Golf-nrt: Integrating global context and local geometry 10 for few-shot view synthesis. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21349– 21359, 2025. 2

2025
[50]

Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 5

2004
[51]

Murf: multi-baseline radiance fields

Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: multi-baseline radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041–20050, 2024. 2, 6

2024
[52]

Vr-nerf: High- fidelity virtualized walkable spaces

Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bul `o, Lorenzo Porzi, Peter Kontschieder, Aljaˇz Boˇziˇc, et al. Vr-nerf: High- fidelity virtualized walkable spaces. InSIGGRAPH Asia 2023 Conference Papers, pages 1–12, 2023. 1

2023
[53]

Neural visibility field for uncertainty-driven active mapping

Shangjie Xue, Jesse Dill, Pranay Mathur, Frank Dellaert, Panagiotis Tsiotra, and Danfei Xu. Neural visibility field for uncertainty-driven active mapping. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18122–18132, 2024. 4

2024
[54]

Contranerf: Gen- eralizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning

Hao Yang, Lanqing Hong, Aoxue Li, Tianyang Hu, Zhen- guo Li, Gim Hee Lee, and Liwei Wang. Contranerf: Gen- eralizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16508–16517, 2023. 1

2023
[55]

Cross-ray neural radiance fields for novel- view synthesis from unconstrained image collections

Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel- view synthesis from unconstrained image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15901–15911, 2023. 2

2023
[56]

Neural radiance field-based visual rendering: A comprehensive review,

Mingyuan Yao, Yukang Huo, Yang Ran, Qingbin Tian, Ruifeng Wang, and Haihua Wang. Neural radiance field- based visual rendering: A comprehensive review.arXiv preprint arXiv:2404.00714, 2024. 1

work page arXiv 2024
[57]

pixelnerf: Neural radiance fields from one or few images

Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4578–4587, 2021. 1

2021
[58]

Supernerf: High-precision 3d reconstruction for large-scale scenes.IEEE Transactions on Geoscience and Remote Sens- ing, 2024

Guangyun Zhang, Chaozhong Xue, and Rongting Zhang. Supernerf: High-precision 3d reconstruction for large-scale scenes.IEEE Transactions on Geoscience and Remote Sens- ing, 2024. 1

2024
[59]

Dynamic scene understanding through object- centric voxelization and neural rendering.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2025

Yanpeng Zhao, Yiwei Hao, Siyu Gao, Yunbo Wang, and Xi- aokang Yang. Dynamic scene understanding through object- centric voxelization and neural rendering.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2025. 2

2025
[60]

Caesarnerf: Calibrated semantic representation for few-shot generalizable neural rendering

Haidong Zhu, Tianyu Ding, Tianyi Chen, Ilya Zharkov, Ram Nevatia, and Luming Liang. Caesarnerf: Calibrated semantic representation for few-shot generalizable neural rendering. InEuropean Conference on Computer Vision, pages 71–89. Springer, 2024. 2 11

2024