Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Improving Novel view synthesis of 360$^\circ$ Scenes in Extremely Sparse Views by Jointly Training Hemisphere Sampled Synthetic Images

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding 50 synthetic upper-hemisphere training views to a four-view 3DGS pipeline improves 360-degree novel view synthesis, with 0.98 dB PSNR and 0.08 SSIM gains over InstantSplat.

desk verdict A plausible pipeline for sparse 360 NVS with a clean ablation story, but the headline gain over InstantSplat is measured on too narrow a benchmark to be trusted as-is. read the letter →

arxiv 2505.19264 v1 pith:2SNHC4KH submitted 2025-05-25 cs.CV

Guangan Chen , Anh Minh Truong , Hanhe Lin , Michiel Vlaminck , Wilfried Philips , Hiep Luong This is my paper · ORCID
classification cs.CV
keywords novelviewsynthesis3DGaussianSplattingextremelysparseviews360-degreescenesDUSt3Rupperhemispheresamplingdiffusion-basedimageenhancementsynthetictraining
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 claims that 360-degree novel view synthesis from only four input photographs can be substantially improved by generating fifty additional synthetic training views in the upper hemisphere of the scene and training a 3D Gaussian Splatting (3DGS) model, a fast explicit 3D scene representation, jointly on the real and synthetic images. The synthetic views are rendered from the dense point cloud produced by the multi-view stereo method DUSt3R from the four reference images, then cleaned by a diffusion-based enhancement model retrained on a large multi-view dataset. On nine real scenes from Mip-NeRF 360 and Tanks&Temples, the method reports gains of 0.98 dB in PSNR, 0.08 in SSIM, and a 0.01 decrease in LPIPS over InstantSplat, the strongest pose-free sparse-view baseline. The practical stakes are high: accurate rendering from a handful of photos would make VR/AR scene capture and 3D reconstruction far cheaper and faster.

What carries the argument

The mechanism is the upper-hemisphere view sampler defined by Eq. 1, which divides the hemisphere into $L$ elevation levels, places cameras according to Fibonacci-sequence counts, parameterizes each pose by radius, azimuth, and elevation, and orients every camera toward the scene center. This sampler is what converts a single sparse point cloud into 50 extra training views spread out in 3D space. The second carrying component is the distance-aware weight $\lambda(\pi_s^j)$ in Eq. 3, which lets synthetic views nearer to reference images contribute more to the joint loss of Eq. 2; the third is the retrained DiffBIR enhancer, which suppresses artifacts in point-cloud-rendered images and is trained with contextual loss to tolerate pose misalignment.

What would settle it

A controlled comparison would settle the central claim: train the same pipeline with the same 50 synthetic views sampled from the lower hemisphere, the full sphere, and the camera trajectory, holding enhancement and weighting fixed. If lower-hemisphere or full-sphere sampling matches or beats the upper-hemisphere strategy on the same nine scenes, the paper's geometric explanation is wrong. A second check is to run the method on a scene where DUSt3R is known to omit a large background region (e.g., the sky or a ceiling); the synthetic views should visibly fail in those regions, and if the gain persists despite that, the proposed mechanism is not the source of the improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that, in a four-view 360-degree setting, the main failure mode of 3D Gaussian Splatting is missing angular coverage rather than missing image count: training only on the reference views overfits them. To fix this, the method densely samples 50 camera poses across five elevation levels of the upper hemisphere, orients them toward the scene center, and renders each sampled pose from the DUSt3R point cloud (Eq. 1). Those point-cloud renders are sent through a retrained diffusion-based enhancer, and the cleaned synthetic images enter the 3DGS training loss together with the reference images, with a distance-aware weight that favors synthetic views close to the reference cameras (Eqs. 2 and 3). The measured result is that this joint training beats the InstantSplat baseline by 0.98 dB PSNR, 0.08 SSIM, and 0.01 LPIPS, and the ablations attribute the gain to both the upper-hemisphere sampling strategy and the enhancement step.

Load-bearing premise

The load-bearing premise is that DUSt3R, applied to only four images, produces a point cloud and camera poses accurate enough that images rendered from synthetic upper-hemisphere poses are faithful to the actual scene after diffusion enhancement; the paper's own results concede that large background areas remain unrepresented in the point cloud, producing empty regions, so this premise is only partially satisfied.

Editorial extensions

If this is right

  • Synthetic views sampled in 3D space around the upper hemisphere outperform views sampled along the camera trajectory by 0.18 dB PSNR, so spatial coverage of the hemisphere is an effective knob for sparse-view 3DGS.
  • Cleaning the point-cloud renders with the retrained diffusion enhancer adds 0.25 dB PSNR on top of the sampling gain, meaning artifact removal in synthetic training data is itself a source of improvement.
  • The full pipeline needs only four reference images and no ground-truth poses, so it offers a practical route to sparse 360-degree capture for VR/AR and scene monitoring.
  • The LPIPS gain over the baseline is only 0.01 while PSNR and SSIM gains are larger, which means the perceived-quality improvement is more modest than the pixel-level one.

Reading between the lines

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

  • If the coverage hypothesis is the real driver, then the benefit should grow with the completeness of the DUSt3R point cloud; the paper's own empty-region artifacts suggest the gain is currently capped by missing background geometry.
  • The upper-hemisphere choice is presumably tuned to scenes with a dominant floor and centered objects; a direct test would be to repeat the same sampling on the full sphere or the lower hemisphere on scenes with prominent ceilings or sky.
  • The recipe used to build the enhancement training set (rendering from DUSt3R point clouds, pairing with reference images, retraining with contextual loss) could be transferred to other scene categories, such as street-level or indoor navigation, without altering the 3DGS training loop.
  • Because all reported numbers sit near 14 dB PSNR, the method is best understood as a strong initialization for extremely sparse reconstruction rather than as a substitute for dense capture.
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

4 major / 7 minor

Summary. The paper proposes a framework for novel view synthesis (NVS) of 360-degree scenes from extremely sparse inputs (four views). It uses DUSt3R to estimate camera poses and generate a dense point cloud, densely samples additional camera poses in the upper hemisphere, renders synthetic images from the point cloud, enhances these images with a retrained DiffBIR model, and finally trains a 3D Gaussian Splatting (3DGS) model jointly on the reference and enhanced synthetic images. Experiments on Mip-NeRF 360 and Tanks & Temples scenes report improvements over InstantSplat, COGS, and DiffusioNeRF, with the central claim being an average improvement of 0.98 dB PSNR, 0.08 SSIM, and 0.01 LPIPS over InstantSplat. The paper also provides ablations isolating the contribution of the synthetic views, perceptual loss, distance-aware weighting, and image enhancement.

Significance. If the reported improvements hold under an unbiased evaluation, the method would be a practical contribution to sparse-view 360-degree NVS: it is simple, builds on open-source components, and the source code is released. The ablation study is a strength, as it demonstrates that each component contributes positively. The evaluation uses standard datasets with held-out test images, which is appropriate for comparing against baselines. However, the current evidence is weakened by the selection protocol: the same nine scenes are used for both hyperparameter tuning and final reporting, the training views are manually chosen for 360-degree coverage, and no error bars are provided. The claimed LPIPS gain is not consistent across scenes, with several scenes showing equal or worse LPIPS than InstantSplat. Additionally, because the synthetic views are rendered from the same DUSt3R point cloud that is derived from the reference views, they do not introduce truly external scene information; the method's success depends on the accuracy of that point cloud, which the paper itself notes is incomplete in large background areas.

major comments (4)
  1. [Table 1, Sec. 3.5] The central claim of an average 0.01 LPIPS improvement over InstantSplat is not supported by the per-scene numbers: on Garden (0.61 vs 0.60), Bicycle (0.64 vs 0.63), and Stump (0.63 vs 0.63), the proposed method is worse or equal, so the average gain is within rounding error and the improvement is not consistent.
  2. [Sec. 3.1, Sec. 3.3, Sec. 3.5] The evaluation protocol does not separate model selection from assessment: the hyperparameters (perceptual losses λP=0.5/0.1, elevation cap τ=0.8, L=5 levels, N=50 synthetic views, 5 DDIM steps, distance-weight normalization Dmax) are fixed using the same nine scenes on which the final numbers are reported, and the four reference views are manually selected for 360-degree coverage with no random subsets or variance estimates. A held-out validation set with multiple view selections and reported mean/std is needed before the improvement can be attributed to the method rather than to tuning and selection.
  3. [Sec. 3.5, Eq. (2)] The paper admits that 'large background areas remain unrepresented in the point cloud, leading to some empty regions'; since the synthetic views in Eq. (2) are rendered from this incomplete point cloud, the added supervision can inject incorrect geometry in those regions. The manuscript provides no analysis of how the results depend on point-cloud coverage or on the number and placement of synthetic views, which is a load-bearing unknown for the claimed benefit.
  4. [Sec. 2.3, Eq. (1)] The upper-hemisphere sampling strategy assumes that a scene has a well-defined center and that cameras point directly at it; the method is only evaluated on scenes selected for a centered object (Sec. 3.1), so it is unclear whether the gains extend to general 360-degree scenes without a centered object, and the title's claim of improving '360° scenes' is broader than the evaluated scope.
minor comments (7)
  1. [Sec. 2.1] There are typos: 'spare view' should be 'sparse view' and 'efficent' should be 'efficient'.
  2. [Sec. 3.3] 'imprically' should be 'empirically'.
  3. [Sec. 3.5] 'apperance' should be 'appearance'.
  4. [Eq. (1)] The Fibonacci sequence F(Tl) is used to determine the number of cameras per elevation level, but Tl is not defined and the mapping from Fibonacci numbers to camera counts is unclear.
  5. [Eq. (3)] Dmax is not explicitly defined; state whether it is the maximum pairwise distance among reference camera poses or the maximum distance of any sampled pose to a reference pose.
  6. [Sec. 3.4, references] The perceptual loss LP is said to follow [17] (Simonyan's VGG paper), but LPIPS is the metric from [27]; cite [27] where the perceptual loss is introduced.
  7. [Table 1] The caption states 'best performance values are highlighted', but no highlighting or bolding appears in the table text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the synthetic-view training signal is derived from DUSt3R point clouds and external multi-view datasets, and Table 1 is measured on held-out test views.

full rationale

This paper's derivation chain is self-contained and does not reduce to its inputs. The pipeline is: DUSt3R estimates camera poses and a dense point cloud from the four reference images (Sec. 2.2); upper-hemisphere poses are sampled by the geometric formula in Eq. (1); point-cloud renders are enhanced by a DiffBIR model retrained on 52,552 pairs from WildRGB-D, MVImgNet, and DL3DV-10K (Sec. 2.4); and 3DGS is trained with the loss in Eqs. (2)-(4), which compares reference images and enhanced synthetic images to 3DGS renders. None of these equations defines the predicted output in terms of itself, and the enhancement model is trained on external datasets, not on the Mip-NeRF 360 or Tanks&Temples evaluation scenes. The quantitative claim in Table 1 compares against InstantSplat, COGS, and DiffusioNeRF using held-out test views registered with iComma, not the synthetic views used in training. The single self-citation, [3], supports only the introductory point that large scenes require reacquisition when changed and is not load-bearing for the method or the evaluation. The admitted empty-region limitation in Sec. 3.5 and the absence of error bars or a separate validation set are experimental robustness concerns, but they are not circularity under the criteria used here.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The pipeline introduces no new physical or mathematical entities. It relies on existing tools (DUSt3R, 3DGS, DiffBIR) and several hand-chosen hyperparameters. The main architectural assumption is that synthetic upper-hemisphere views rendered from a DUSt3R point cloud are a reliable source of additional supervision, which is only partially supported by the paper's own failure cases.

free parameters (5)
  • Number of synthetic views N = 50
    Set for all scenes in Sec. 3.3; no sensitivity analysis shown, and the ablation in Table 3 compares only 50 vs 50.
  • Elevation levels L = 5
    The upper hemisphere is divided into five elevation levels (Sec. 3.3). This choice affects sampling density and is not varied.
  • Maximum elevation ratio tau = 0.8
    Described as set 'imprically' in Sec. 3.3; it bounds the elevation angle in Eq. 1 and controls how high the sampled views go.
  • Perceptual loss weight lambda_P for synthetic images = 0.1
    Set to 0.1 for synthetic views and 0.5 for reference views (Sec. 3.3); per-component contribution is shown in the Table 2 ablation, but no validation split is used.
  • Distance-weight normalization Dmax = Maximum distance among reference camera poses
    Normalizes the distance-aware weight lambda(pi) in Eq. 3, borrowed from GaussianObject; it is a heuristic with no independent justification.
assumptions (5)
  • domain assumption DUSt3R provides sufficiently accurate camera poses and a dense point cloud for 360-degree scenes from four input views.
    Invoked in Sec. 2.2 as the foundation for pose estimation and point-cloud rendering. The paper itself acknowledges incomplete background regions, so this assumption is only partially satisfied.
  • domain assumption The test scenes are suitable for the centered-object setup, and the four selected views cover the 360-degree range.
    Sec. 3.1 explicitly selects only scenes 'with a centered object only' and manually ensures angular and spatial diversity, which biases the evaluation toward favorable cases.
  • domain assumption The retrained DiffBIR enhancement removes point-cloud rendering artifacts without introducing new content into the images.
    Stated in Sec. 2.4 and relied on during training, but no quantitative check for hallucinated content is provided, despite the generative nature of diffusion models.
  • ad hoc to paper The distance-aware weight lambda(pi) from Eq. 3 is a valid weighting for synthetic view loss.
    Borrowed from GaussianObject [25]; the ablation in Table 2 shows a joint benefit with perceptual loss, but does not isolate the weighting scheme itself.
  • domain assumption 3DGS trained for 1000 iterations with the combined reference and synthetic images converges to a useful scene representation.
    Standard 3DGS training assumptions apply; the 1000-iteration budget is taken from InstantSplat and not independently validated for this combined training set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Novel view synthesis of 360$^\circ$ Scenes in Extremely Sparse Views by Jointly Training Hemisphere Sampled Synthetic Images." pith.science (2026). https://pith.science/paper/2SNHC4KH

@misc{pith2026250519264,
  author       = {Pith},
  title        = {Pith review of: Improving Novel view synthesis of 360$^\circ$ Scenes in Extremely Sparse Views by Jointly Training Hemisphere Sampled Synthetic Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SNHC4KH}},
  note         = {Machine review of arXiv:2505.19264}
}
abstract

Novel view synthesis in 360$^\circ$ scenes from extremely sparse input views is essential for applications like virtual reality and augmented reality. This paper presents a novel framework for novel view synthesis in extremely sparse-view cases. As typical structure-from-motion methods are unable to estimate camera poses in extremely sparse-view cases, we apply DUSt3R to estimate camera poses and generate a dense point cloud. Using the poses of estimated cameras, we densely sample additional views from the upper hemisphere space of the scenes, from which we render synthetic images together with the point cloud. Training 3D Gaussian Splatting model on a combination of reference images from sparse views and densely sampled synthetic images allows a larger scene coverage in 3D space, addressing the overfitting challenge due to the limited input in sparse-view cases. Retraining a diffusion-based image enhancement model on our created dataset, we further improve the quality of the point-cloud-rendered images by removing artifacts. We compare our framework with benchmark methods in cases of only four input views, demonstrating significant improvement in novel view synthesis under extremely sparse-view conditions for 360$^\circ$ scenes.

Discussion (0). Continue with ORCID 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. Sparse-View 3D Reconstruction: Recent Advances and Open Challenges

    cs.CV 2025-07 conditional novelty 3.0 of 10

    A comprehensive survey that organizes sparse-view 3D reconstruction methods into geometry-based, NeRF, 3DGS, and diffusion-based categories, with benchmarks and open challenges.

Reference graph

Works this paper leans on

32 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    INTRODUCTION Novel view synthesis (NVS) is a key challenge in com- puter vision, focusing on rendering images from previ- ously unobserved viewpoints. Recent advancements, such as Neural Radiance Fields (NeRF) [12] and 3D Gaussian Splatting (3DGS) [6], have demonstrated remarkable ca- pabilities in generating novel views using dense captured images of a s...

    arXiv 2025

  2. [2]

    Preliminary 3DGS [6] is an explicit representation for modeling scenes with parameterized 3D Gaussians

    METHOD 2.1. Preliminary 3DGS [6] is an explicit representation for modeling scenes with parameterized 3D Gaussians. A 3D Gaussian is param- eterized by G = {µ, q, s, o, c, sh}, where µ is spatial mean, q is rotation quaternion, s is scaling vector, o is opacity, c is a view-dependent color, and sh is spherical harmonic (SH) coefficients. A 3D scene is rep...

    work page

  3. [3]

    Dataset The two real-world benchmarking datasets featuring large- scale scenes: Mip-NeRF 360 [1], and Tanks&Temples ( 360◦ scenes) [7] are used to evaluate our method

    EXPERIMENTS 3.1. Dataset The two real-world benchmarking datasets featuring large- scale scenes: Mip-NeRF 360 [1], and Tanks&Temples ( 360◦ scenes) [7] are used to evaluate our method. Specifically, we evaluate our method in four-view cases, i.e., M = 4. To re- construct a 360◦ scene, it must have overlap between sampled images, even in extremely spare vi...

    work page

  4. [4]

    The framework uses reference and synthetic images to train a 3DGS model

    CONCLUSION AND FUTURE WORK We propose a framework to improve the performance of 3DGS for 360 ◦ scenes with extremely sparse views. The framework uses reference and synthetic images to train a 3DGS model. Those synthetic images are created with cam- era poses sampled in the upper hemisphere space of scenes and DUSt3R-created point clouds. Moreover, we remo...

    work page

  5. [5]

    ACKNOWLEDGMENT This research received partial funding from the Flemish Gov- ernment under the ‘Onderzoeksprogramma Artificiele Intelli- gentie (AI) Vlaanderen’ programme

    work page

  6. [6]

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022

    work page 2022

  7. [7]

    Brooks, A

    T. Brooks, A. Holynski, and A. A. Efros. InstructPix2Pix: Learning To Follow Image Editing Instructions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 18392–18402, 2023

    work page 2023

  8. [8]

    G. Chen, M. Vlaminck, W. Philips, and H. Q. Luong. Multi-View 3D Reconstruction for Construction Site Monitoring. In Conference on VISAPP, 2024

    work page 2024

Show all 32 references
  1. [9]

    Z. Fan, W. Cong, K. Wen, K. Wang, J. Zhang, X. Ding, D. Xu, B. Ivanovic, M. Pavone, G. Pavlakos, et al. InstantSplat: Unbounded Sparse-view Pose-free Gaussian Splatting in 40 Seconds. arXiv preprint arXiv:2403.20309, 2024

    2024 arXiv

  2. [10]

    Jiang, Y

    K. Jiang, Y . Fu, M. Varma T, Y . Belhe, X. Wang, H. Su, and R. Ra- mamoorthi. A Construct-Optimize Approach to Sparse View Syn- thesis without Camera Pose. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024

    2024

  3. [11]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics, 42(4):139–1, 2023

    2023

  4. [12]

    Knapitsch, J

    A. Knapitsch, J. Park, Q.-Y . Zhou, and V . Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1–13, 2017

    2017

  5. [13]

    Liang, J

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte. SwinIR: Image Restoration Using Swin Transformer. In IEEE Inter- national Conference on Computer Vision, pages 1833–1844, 2021

    2021

  6. [14]

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, W. Ouyang, Y . Qiao, and C. Dong. DiffBIR: Toward Blind Image Restoration with Generative Diffusion Prior. arXiv preprint arXiv:2308.15070, 2023

    2023 arXiv

  7. [15]

    L. Ling, Y . Sheng, Z. Tu, W. Zhao, C. Xin, K. Wan, L. Yu, Q. Guo, Z. Yu, Y . Lu, et al. DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision. In IEEE Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024

    2024

  8. [16]

    Mechrez, I

    R. Mechrez, I. Talmi, and L. Zelnik-Manor. The Contextual Loss for Image Transformation with Non-Aligned Data. In IEEE Conference on European Conference on Computer Vision, pages 768–783, 2018

    2018

  9. [17]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021

    2021

  10. [18]

    S. Paul, C. Wewer, B. Schiele, and J. E. Lenssen. Sp2360: Sparse- view 360◦ Scene Reconstruction using Cascaded 2D Diffusion Priors. In ECCV 2024 Workshop on Wild 3D: 3D Modeling, Reconstruction, and Generation in the Wild , 2024

    2024

  11. [19]

    Poole, A

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall. DreamFusion: Text- to-3D using 2D Diffusion. In International Conference on Learning Representations, 2023

    2023

  12. [20]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-Resolution Image Synthesis With Latent Diffusion Models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022

    2022

  13. [21]

    J. L. Schonberger and J.-M. Frahm. Structure-From-Motion Revis- ited. In IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 4104–4113, 2016

    2016

  14. [22]

    Simonyan

    K. Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  15. [23]

    Y . Sun, X. Wang, Y . Zhang, J. Zhang, C. Jiang, Y . Guo, and F. Wang. iComMa: Inverting 3D Gaussian Splatting for Camera Pose Estima- tion via Comparing and Matching. arXiv preprint arXiv:2312.09031, 2023

    2023 arXiv

  16. [24]

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud. DUSt3R: Geometric 3D Vision Made Easy. In IEEE Conference on Computer Vision and Pattern Recognition, pages 20697–20709, 2024

    2024

  17. [25]

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, 2004

    2004

  18. [26]

    R. Wu, B. Mildenhall, P. Henzler, K. Park, R. Gao, D. Watson, P. P. Srinivasan, D. Verbin, J. T. Barron, B. Poole, et al. ReconFusion: 3D Reconstruction with Diffusion Priors. In IEEE Conference on Com- puter Vision and Pattern Recognition, pages 21551–21561, 2024

    2024

  19. [27]

    Wynn and D

    J. Wynn and D. Turmukhambetov. DiffusioNeRF: Regularizing Neu- ral Radiance Fields With Denoising Diffusion Models. In IEEE Con- ference on Computer Vision and Pattern Recognition , pages 4180– 4189, 2023

    2023

  20. [28]

    H. Xia, Y . Fu, S. Liu, and X. Wang. RGBD Objects in the Wild: Scaling Real-World 3D Object Learning from RGB-D Videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 22378–22389, 2024

    2024

  21. [29]

    Xiong, S

    H. Xiong, S. Muttukuru, R. Upadhyay, P. Chari, and A. Kadambi. SparseGS: Real-Time 360◦ Sparse View Synthesis using Gaussian Splatting. arXiv preprint arXiv:2312.00206, 2023

    2023 arXiv

  22. [30]

    C. Yang, S. Li, J. Fang, R. Liang, L. Xie, X. Zhang, W. Shen, and Q. Tian. GaussianObject: High-Quality 3D Object Recon- struction from Four Views with Gaussian Splatting. arXiv preprint arXiv:2402.10259, 2024

    2024 arXiv

  23. [31]

    X. Yu, M. Xu, Y . Zhang, H. Liu, C. Ye, Y . Wu, Z. Yan, C. Zhu, Z. Xiong, T. Liang, et al. MVImgNet: A Large-Scale Dataset of Multi- View Images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9150–9161, 2023

    2023

  24. [32]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The Un- reasonable Effectiveness of Deep Features as a Perceptual Metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018

    2018

Pith tools

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