Pith. sign in

REVIEW 4 major objections 4 minor 26 references

ExtraGS: Enhancing Endoscopic View Extrapolation via Diffusion-Guided 3D Gaussian Splatting

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read ExtraGS claims that a diffusion-guided 3D Gaussian Splatting loop can render artifact-free endoscopic views far outside the recorded camera trajectory.

desk verdict The SOTA claim collapses on LPIPS and PSNR, but the endoscopic-specific sampling and confidence weighting are real contributions worth a careful revision. read the letter →

arxiv 2607.12785 v2 pith:OPU5KSTZ submitted 2026-07-14 cs.CV

Cheng-Tai Hsieh , Jiwei Shan , Han Fang , Jianshu Hu , Tao Ni , Lijun Han , Yutong Ban , Shing Shin Cheng
show 1 more author
Hesheng Wang
This is my paper · ORCID
classification cs.CV
keywords 3DGaussianSplattingnovelviewsynthesisextrapolationendoscopydiffusionpriorpseudo-observationsuncertainty-guidedsamplingsurgicalscenereconstruction
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

ExtraGS targets a known failure of neural rendering in endoscopy: when a 3D Gaussian Splatting model trained on a sparse video is asked to render a view far from the recorded path, it produces holes, floaters, and distorted tissue because those regions were never observed. The paper's proposed fix is to treat the problem as one of missing supervision and supply it with generated content. A video diffusion model, conditioned on the nearest real frame, turns coarse renderings of blind spots into plausible pseudo-observations, which then fine-tune the 3D scene under a confidence weighting that preserves reliably reconstructed areas. The paper argues this loop—explore uncertain regions, render, refine with diffusion, fine-tune with confidence weights—reduces extrapolation artifacts and outperforms prior NeRF and 3DGS baselines on colonoscopic datasets. A sympathetic reader would care because the capability to synthesize plausible off-trajectory views could expand an endoscope's effective field of view without new hardware.

What carries the argument

The load-bearing mechanism is the pseudo-observation loop: a conditional video-diffusion denoiser turns a coarse rendering from a virtual camera into a refined image, which is used as pseudo-ground truth in a second optimization pass. The loop is protected by two confidence mechanisms: an image-level weight derived from the Gaussian IoU between the virtual and nearest real view, and a pixel-level weight derived from the LPIPS distance between the coarse and refined renderings. The final loss combines a real-view anchor term and a confidence-weighted virtual-view term, which is what prevents the generated content from corrupting well-observed geometry.

What would settle it

Record a colonoscopy with two synchronized cameras, train ExtraGS on one camera only, and compare its synthesized views from the second camera's viewpoints against the actual footage; systematic divergence in the regions that were unobserved by the first camera would show the pseudo-observations are not faithful supervision.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that large-baseline view extrapolation in endoscopic scenes can be made artifact-free by a three-stage loop. First, a coarse 3D Gaussian Splatting scene is optimized from posed RGB-D frames. Second, virtual cameras are actively sampled toward uncertain, dark, and under-observed regions using an illumination-aware coverage gain. Third, rendered views are enhanced by a conditional video-diffusion prior into pseudo-observations, and the Gaussian scene is fine-tuned with an image-level coverage weight (based on Gaussian IoU) and a pixel-level perceptual-difference weight, so that only diffusion-corrected content is emphasized. The paper reports that

Load-bearing premise

The diffusion model's pseudo-observations are anatomically correct enough to supervise regions no real camera ever saw; if it hallucinates plausible but wrong tissue, the fine-tuning step bakes those errors into the 3D model.

Editorial extensions

If this is right

  • If correct, ExtraGS would let a 3DGS scene trained on a thin endoscopic trajectory render views from substantially different angles, making 'looking around' possible without additional camera hardware.
  • Because the confidence weights keep diffusion edits local, the method could be applied on top of any existing 3DGS reconstruction without discarding the original observations.
  • The finding that dark regions are the priority for exploration suggests that lighting-aware sampling is a transferable design for other narrow-FOV reconstruction tasks.
  • The static-scene assumption means the current result applies to offline reconstruction; extending to moving tissue would require deformation-aware conditioning.

Reading between the lines

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

  • Editorial inference: if the diffusion prior ever generates anatomically plausible but factually wrong tissue, the fine-tuned model will appear smooth while being clinically misleading; the paper does not check synthesized blind spots against ground truth, so a validation with a second camera is the natural next step.
  • Editorial inference: the pseudo-observation loop is modular; any generative model that can supply plausible pixels could replace the video diffusion prior, so the framework's value may track improvements in video generation.
  • Editorial inference: the G-IOU/LPIPS confidence scheme could be reused for other self-supervised 3D refinement tasks, e.g., filling occlusion holes in stereo reconstructions, independent of endoscopy.
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 / 4 minor

Summary. The paper proposes ExtraGS, a framework for endoscopic novel-view extrapolation that combines 3D Gaussian Splatting (3DGS) with uncertainty-guided virtual camera sampling and video-diffusion pseudo-observations. Starting from a coarse 3DGS reconstruction, the method estimates a feasible free-space region, samples virtual trajectories toward under-observed areas, refines the rendered views with a video diffusion model (DynamiCrafter), and jointly fine-tunes the 3DGS using the real views plus confidence-weighted pseudo-observations. The authors claim state-of-the-art performance on the C3VDv2 dataset, with code released.

Significance. If the central claim were supported, the paper would be a relevant contribution to endoscopic view extrapolation: combining diffusion priors with 3DGS and actively sampling virtual viewpoints is a natural and potentially practical direction, and the release of code is a strength. However, the quantitative evidence in the paper's own Table I contradicts the 'state-of-the-art' claim: ExtraGS has substantially worse LPIPS than UCNeRF on both scenarios and worse PSNR on the distanced scenario. The paper also lacks error bars and does not validate the diffusion-generated content in extrapolated regions against ground truth. These are load-bearing issues, not presentation details.

major comments (4)
  1. [Abstract and §IV-B1, Table I] The abstract claims that ExtraGS 'achieves state-of-the-art performance in endoscopic novel view synthesis,' but Table I contradicts this. On Sparse: Cecum t1, ExtraGS reports LPIPS 0.483 vs UCNeRF 0.244; on Distanced: Ascending t2, ExtraGS reports LPIPS 0.443 vs 0.214 and PSNR 18.75 vs UCNeRF 20.15. Only SSIM is consistently better. No metric-specific definition of 'state-of-the-art' is given, and there are no error bars, so even the SSIM advantage cannot be assessed for significance.
  2. [Table I vs Table II] The same setting (C3VDv2 Cecum t1) is reported with different numbers for the full model: Table I gives PSNR 20.23 / SSIM 0.769, while Table II gives PSNR 20.38 / SSIM 0.770. This unresolved inconsistency suggests unstated differences in evaluation protocol or data split, which undermines the reliability of both the main comparison and the ablation study.
  3. [Sec. III-E/III-F, Eq. (18)] The framework uses diffusion-generated pseudo-observations as supervision for previously unobserved regions, but no ground-truth check of those extrapolated regions is performed. If the video diffusion prior hallucinates plausible but anatomically wrong tissue, the confidence-weighted fine-tuning step will bake those errors into the 3D model. The paper's own conclusion states that the results 'demonstrate potential rather than validated intraoperative reliability,' which is not sufficient support for the headline claim. Held-out validation of extrapolated viewpoints is needed.
  4. [§IV-C, Table II] The ablation removes only the illumination- and fold-aware components. It does not ablate the diffusion prior, the uncertainty-guided sampling, or the confidence-weighted fine-tuning, which are the core contributions. The reported full-model numbers also differ from Table I. As presented, the ablation does not provide evidence for the central mechanism of the method.
minor comments (4)
  1. [Abstract/Introduction] The phrase 'state-of-the-art' should be replaced with metric-specific and scenario-specific claims, e.g., 'improves SSIM under both sparse and distanced settings, but with worse LPIPS.'
  2. [§IV-A3] Key hyperparameters (N_Tr=20, S=64, D=32, τ=0.5, top-3 selection, 15K fine-tune iterations) are set without sensitivity analysis. Given that several of these are ad hoc, the robustness of the method is not demonstrated.
  3. [Fig. 3] The qualitative superiority is asserted from small images. Enlarged crops and per-method zoom-ins would help the reader verify the claims about fine vessels and specular highlights.
  4. [§IV-B1] The text says ExtraGS 'comprehensively outperforms' baselines, but the table shows it loses on LPIPS in both settings. The wording should be corrected to accurately reflect the reported numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ExtraGS's pseudo-observations come from an external diffusion prior, and no claimed prediction is defined by construction in terms of its inputs or a self-citation chain.

full rationale

The derivation chain is externally grounded. The paper first trains a coarse 3DGS model (Eq. 3), then samples virtual cameras using coverage and illumination heuristics (Eqs. 8-16), renders coarse views (Eq. 17), enhances them with a pretrained video diffusion backbone DynamiCrafter (Eq. 18), and finally fine-tunes with confidence-weighted losses (Eqs. 24-25). The pseudo-observations I_P(v) are not defined in terms of the evaluation targets; they are generated by a pretrained diffusion model from a coarse rendering plus a real anchor view, and the test metrics are computed against real ground-truth images, not against I_P. No fitted constant or network weight is set using the held-out test views. The reuse of ExploreGS and 3DGS-Enhancer components is standard external prior work by other authors, and no self-citation is invoked as the load-bearing justification for the central optimization loop. The paper's own limitation statement - 'the present results demonstrate potential rather than validated intraoperative reliability' - is an honest validity caveat, not a circularity signal. The unsupported 'state-of-the-art' claim relative to Table I is a correctness/evidence concern, outside the circularity audit as defined here.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities; pseudo-observations are generated images, not independent entities. The load-bearing assumptions are the correctness of the diffusion prior and the reliability of the coarse 3DGS-derived occupancy grid, both of which are unvalidated for surgical blind spots.

free parameters (10)
  • Occupancy threshold τ = 0.5
    Sec III-C: voxels with visibility below 0.5 are marked occupied; this controls the feasible region for all virtual camera sampling.
  • Candidate motion primitives A = 14
    Sec III-D1: six translations + four rotations + four orbital moves; defines the search action space.
  • Virtual trajectory count N_Tr = 20
    Sec IV-A3: number of sampled virtual trajectories; no sensitivity analysis is provided.
  • Occupancy grid resolution S = 64
    Sec IV-A3: resolution of the occupancy grid; affects feasible-region granularity.
  • Viewing-direction bins D = 32
    Sec IV-A3: discretization for the coverage map; affects information-gain estimates.
  • Top-k candidate selection = 3
    Sec IV-A3: top-3 candidates expanded per step.
  • Rotation penalty threshold τ_R and factor η = not reported
    Sec III-D4: fold-aware pose control parameters introduced but values are never given.
  • Illumination weighting λ_l = not reported
    Sec III-D3: weight on cosine similarity toward darkest direction; value absent.
  • Fine-tune iterations / densification horizon = 15K / 9K
    Sec IV-A3: joint fine-tuning schedule; chosen by hand.
  • GPT-4o caption prompt = not specified
    Sec IV-A3: text conditioning for DynamiCrafter is generated by GPT-4o, but the prompts are not released, blocking exact reproduction.
assumptions (6)
  • standard math 3DGS splatting and diffusion forward/reverse process equations (Eqs. 1–5) are correct
    Used as building blocks in Sec III-B; not derived by this paper.
  • domain assumption The scene is approximately static during reconstruction
    Sec V: the paper admits it 'assumes relatively static scenes' while real surgery involves tissue deformation.
  • domain assumption Camera poses and depth maps are accurate
    Sec III-A and Fig. 2: pipeline consumes posed RGB+depth with poses estimated via COLMAP; pose error would propagate into the coarse 3DGS and virtual sampling.
  • domain assumption Diffusion-enhanced pseudo-observations are anatomically plausible in unobserved regions
    Sec III-E / Eq. 18: pseudo-GT is generated by DynamiCrafter without ground truth for blind spots; the conclusion says results show 'potential rather than validated intraoperative reliability'.
  • domain assumption Occupancy grid from coarse 3DGS reliably separates free space from anatomy
    Sec III-C: virtual cameras are restricted to voxels with visibility ≥ 0.5; if the coarse reconstruction is wrong, exploration is misdirected.
  • domain assumption Illumination-derived uncertainty and fold-aware smoothness improve extrapolation coverage
    Sec III-D: the heuristics that darkness correlates with under-reconstruction and that smooth camera motion is safer are plausible but tested by only one combined ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ExtraGS: Enhancing Endoscopic View Extrapolation via Diffusion-Guided 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/OPU5KSTZ

@misc{pith2026260712785,
  author       = {Pith},
  title        = {Pith review of: ExtraGS: Enhancing Endoscopic View Extrapolation via Diffusion-Guided 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OPU5KSTZ}},
  note         = {Machine review of arXiv:2607.12785}
}
read the original abstract

Robot-assisted minimally invasive surgery (MIS) critically depends on reliable endoscopic perception for navigation and safety. However, conventional endoscopes provide only a limited field of view, leaving large portions of the surrounding anatomy unobserved. Recent neural rendering approaches, such as Neural Radiance Fields and 3D Gaussian Splatting, enable novel view synthesis from endoscopic videos, but their reliance on sparse observations often leads to severe artifacts when extrapolating beyond the training trajectory. In this work, we propose ExtraGS, a framework for enhancing endoscopic view extrapolation through diffusion-guided 3D Gaussian Splatting. Starting from an initial reconstruction, we introduce an uncertainty-guided virtual camera sampling strategy to actively explore blind spots and maximize information gain. The rendered views from these sampled locations are refined using a diffusion model to recover plausible anatomical structures, producing pseudo-observations that guide further optimization. To prevent the generated content from degrading reliable regions, we adopt a confidence-weighted fine-tuning strategy when incorporating these pseudo-observations. Extensive experiments on multiple public endoscopic datasets demonstrate that ExtraGS significantly reduces extrapolation artifacts and achieves state-of-the-art performance in endoscopic novel view synthesis.

Figures

Figures reproduced from arXiv: 2607.12785 by the authors.

Figure 1
Figure 1. Extra-GS enables exploration-driven reconstruction. Top: Cross￾sectional view of the dataset mesh illustrating training (red) and testing (blue) camera trajectories. Bottom: Representative renderings from off￾trajectory viewpoints. While standard 3DGS suffers from severe artifacts when deviating from the training path, our method maintains structural completeness and high fidelity under large viewpoint shifts. 3D an… view at source ↗
Figure 2
Figure 2. Overview of the Extra-GS framework. Our proposed pipeline consists of three primary stages: (a) Scene Initialization: A baseline 3D Gaussian Splatting (3DGS) representation is constructed utilizing available endoscopic RGB and depth images, alongside camera poses estimated via COLMAP. (b) Pseudo-Observation Generation: Novel viewpoints are synthesized through strategic virtual camera sampling and subsequently enhanc… view at source ↗
Figure 3
Figure 3. Qualitative comparison of novel view synthesis on the C3VD dataset. Rendered novel views from baseline methods (3DGS, UCNeRF, GaussianPancakes) and our proposed ExtraGS are shown alongside the Ground Truth. The visualizations illustrate the rendering performance of each approach in reconstructing challenging endoscopic environments characterized by complex tissue topologies, fine vascular networks, and specular high… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

  1. [1]

    Guidelines for colonoscopy surveillance after screening and polypectomy: a consensus update by the us multi- society task force on colorectal cancer,

    D. A. Lieberman, D. K. Rex, S. J. Winawer, F. M. Giardiello, D. A. Johnson, and T. R. Levin, “Guidelines for colonoscopy surveillance after screening and polypectomy: a consensus update by the us multi- society task force on colorectal cancer,”Gastroenterology, vol. 143, no. 3, pp. 844–857, 2012

    2012

  2. [2]

    How commonly is upper gastrointestinal cancer missed at endoscopy? a meta-analysis,

    S. Menon and N. Trudgill, “How commonly is upper gastrointestinal cancer missed at endoscopy? a meta-analysis,”Endoscopy interna- tional open, vol. 2, no. 02, pp. E46–E50, 2014

    2014

  3. [3]

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,

    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,” inComputer Vision – ECCV 2020, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 405–421

    2020

  4. [4]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Trans. Graph., vol. 42, no. 4, July 2023. [Online]. Available: https://doi.org/10.1145/3592433

    doi:10.1145/3592433 2023

  5. [5]

    Neural rendering for stereo 3d reconstruction of deformable tissues in robotic surgery,

    Y . Wang, Y . Long, S. H. Fan, and Q. Dou, “Neural rendering for stereo 3d reconstruction of deformable tissues in robotic surgery,” in International conference on medical image computing and computer- assisted intervention. Springer, 2022, pp. 431–441

    2022

  6. [6]

    Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos,

    R. Zha, X. Cheng, H. Li, M. Harandi, and Z. Ge, “Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos,” inInternational conference on medical image computing and computer-assisted intervention. Springer, 2023, pp. 13–23

    2023

  7. [7]

    Gaussian pancakes: geometrically-regularized 3d gaus- sian splatting for realistic endoscopic reconstruction,

    S. Bonilla, S. Zhang, D. Psychogyios, D. Stoyanov, F. Vasconcelos, and S. Bano, “Gaussian pancakes: geometrically-regularized 3d gaus- sian splatting for realistic endoscopic reconstruction,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 274–283

    2024

  8. [8]

    Deformable gaussian splatting for efficient and high-fidelity recon- struction of surgical scenes,

    J. Shan, Z. Cai, C.-T. Hsieh, L. Han, S. S. Cheng, and H. Wang, “Deformable gaussian splatting for efficient and high-fidelity recon- struction of surgical scenes,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 10 545–10 551

    2025

Show all 26 references
  1. [9]

    Uw-dnerf: Deformable soft tissue reconstruction with uncertainty- guided depth supervision and local information integration,

    J. Shan, Z. Zhang, H. Li, C.-T. Hsieh, Y . Li, W. Wu, and H. Wang, “Uw-dnerf: Deformable soft tissue reconstruction with uncertainty- guided depth supervision and local information integration,”IEEE Transactions on Medical Imaging, 2025

    2025

  2. [10]

    Lightneus: Neural surface reconstruction in endoscopy using illumination de- cline,

    V . M. Batlle, J. M. Montiel, P. Fua, and J. D. Tard ´os, “Lightneus: Neural surface reconstruction in endoscopy using illumination de- cline,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023, pp. 502–512

    2023

  3. [11]

    4d monocular surgical reconstruction under arbitrary camera motions,

    J. Shan, Z. Cai, C.-T. Hsieh, Y . Li, H. Liu, L. Han, H. Wang, and S. S. Cheng, “4d monocular surgical reconstruction under arbitrary camera motions,”Medical Image Analysis, p. 103989, 2026

    2026

  4. [12]

    Enerf-slam: A dense endo- scopic slam with neural implicit representation,

    J. Shan, Y . Li, T. Xie, and H. Wang, “Enerf-slam: A dense endo- scopic slam with neural implicit representation,”IEEE Transactions on Medical Robotics and Bionics, vol. 6, no. 3, pp. 1030–1041, 2024

    2024

  5. [13]

    Dds-slam: Dense semantic neural slam for deformable endoscopic scenes,

    J. Shan, Y . Li, L. Yang, Q. Feng, L. Han, and H. Wang, “Dds-slam: Dense semantic neural slam for deformable endoscopic scenes,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 10 837–10 842

    2024

  6. [14]

    Endoflow-slam: Real-time endoscopic slam with flow-constrained gaussian splatting,

    T. Wu, Y . Miao, Z. Li, H. Zhao, K. Dang, J. Su, L. Yu, and H. Li, “Endoflow-slam: Real-time endoscopic slam with flow-constrained gaussian splatting,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 202–212

    2025

  7. [15]

    Nrgs-slam: Monocular non-rigid slam for en- doscopy via deformation-aware 3d gaussian splatting,

    J. Shan, Z. Cai, Y . Li, Y . Chen, L. Han, Y .-h. Liu, H. Wang, and S. S. Cheng, “Nrgs-slam: Monocular non-rigid slam for en- doscopy via deformation-aware 3d gaussian splatting,”arXiv preprint arXiv:2602.17182, 2026

    2026

  8. [16]

    Uc-nerf: Uncertainty-aware conditional neural radiance fields from endoscopic sparse views,

    J. Guo, J. Wang, R. Wei, D. Kang, Q. Dou, and Y .-H. Liu, “Uc-nerf: Uncertainty-aware conditional neural radiance fields from endoscopic sparse views,”IEEE Transactions on Medical Imaging, vol. 44, no. 3, pp. 1284–1296, 2024

    2024

  9. [17]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  10. [18]

    Endo-4dgs: Endoscopic monocular scene reconstruction with 4d gaussian splatting,

    Y . Huang, B. Cui, L. Bai, Z. Guo, M. Xu, M. Islam, and H. Ren, “Endo-4dgs: Endoscopic monocular scene reconstruction with 4d gaussian splatting,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 197–207

    2024

  11. [19]

    Dreamfusion: Text- to-3d using 2d diffusion,

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall, “Dreamfusion: Text- to-3d using 2d diffusion,” inThe Eleventh International Conference on Learning Representations

  12. [20]

    Reconfusion: 3d reconstruction with diffusion priors,

    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,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21 5...

    2024

  13. [21]

    Extranerf: Visibility-aware view extrapolation of neural radiance fields with diffusion models,

    M.-L. Shih, W.-C. Ma, L. Boyice, A. Holynski, F. Cole, B. Curless, and J. Kontkanen, “Extranerf: Visibility-aware view extrapolation of neural radiance fields with diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp....

    2024

  14. [22]

    Exploregs: Explorable 3d scene reconstruction with virtual camera samplings and diffusion priors,

    M. Kim, S. Jeon, I. Cho, M. Yoo, and S. J. Kim, “Exploregs: Explorable 3d scene reconstruction with virtual camera samplings and diffusion priors,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 27 042–27 051

    2025

  15. [23]

    Ewa splatting,

    M. Zwicker, H. Pfister, J. Van Baar, and M. Gross, “Ewa splatting,” IEEE Transactions on Visualization and Computer Graphics, vol. 8, no. 3, pp. 223–238, 2002

    2002

  16. [24]

    3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors,

    X. Liu, C. Zhou, and S. Huang, “3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 133 305–133 327. [Online]. Available: https://proceedings.neurips.cc/pape...

    2024

  17. [25]

    C3vdv2–colonoscopy 3d video dataset with enhanced realism,

    M. V . Golhar, L. S. G. Fretes, L. Ayers, V . S. Akshintala, T. L. Bobrow, and N. J. Durr, “C3vdv2–colonoscopy 3d video dataset with enhanced realism,”arXiv preprint arXiv:2506.24074, 2025

    2025 arXiv

  18. [26]

    Dynamicrafter: Animating open-domain images with video diffusion priors,

    J. Xing, M. Xia, Y . Zhang, H. Chen, W. Yu, H. Liu, G. Liu, X. Wang, Y . Shan, and T.-T. Wong, “Dynamicrafter: Animating open-domain images with video diffusion priors,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 399–417

    2024

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.