Pith. sign in

REVIEW 4 major objections 5 minor 40 references

4DSloMo: 4D Reconstruction for High Speed Scene with Asynchronous Capture

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Ordinary 25 FPS cameras, staggered in start time, reconstruct fast scenes at an effective 100-200 FPS after a video-diffusion model cleans the sparse-view artifacts.

desk verdict A smart repackaging of staggered capture for 4D Gaussian Splatting with a genuinely helpful video-diffusion artifact-fix, but the evidence overstates the frame-rate gain and needs a serious referee. read the letter →

arxiv 2507.05163 v2 pith:HM5WQZ7L submitted 2025-07-07 cs.CV

Yutian Chen , Shi Guo , Tianshuo Yang , Lihe Ding , Xiuyuan Yu , Jinwei Gu , Tianfan Xue This is my paper
classification cs.CV
keywords 4Dreconstructionasynchronouscapturehigh-speedsceneGaussiansplattingvideodiffusionmodelsparse-viewmulti-viewtemporalframe-rateenhancement
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 a multi-view camera array running at ordinary 25 FPS can capture and reconstruct fast motion at an effective 100-200 FPS if the cameras start at deliberately different times. Grouping the cameras and staggering their start instants within each frame interval gives the array temporally denser samples, and the authors show this recovers large, nonlinear motion that synchronous capture misses. Because each timestamp then has fewer viewpoints, the reconstruction develops sparse-view artifacts; the paper trains a video-diffusion model to remove those artifacts while keeping rendering temporally consistent. On temporally subsampled public datasets the approach improves over the underlying 4D Gaussian splatting and over frame-interpolation baselines, and a new 12-sequence real-world asynchronous dataset demonstrates the capture system outside simulation.

What carries the argument

The argument rests on two coupled mechanisms. First, the asynchronous capture scheme places cameras into groups of size $K$ and starts camera $i$ at $t_i = i\cdot(\tau/K) + j\cdot\tau$ within each base interval $\tau$, so $K$ cameras jointly sample $K$ distinct instants per frame and raise the effective frame rate by a factor of $K$. Second, the artifact-fix video diffusion model, built by fine-tuning a large pretrained text-to-video diffusion model with LoRA on 750 noisy-clean pairs rendered from GS4D, maps a rendered video full of floaters to a clean video; the clean video then supervises the 4D Gaussian refinement through an $\ell^1$ plus LPIPS loss. The 4D Gaussian representation itself, with a temporal coordinate and 4D spherical harmonics, supplies the continuous spatiotemporal field that the denser samples constrain.

What would settle it

Record one fast, repetitive scene with both a synchronized 12-camera 25 FPS array and a staggered array, and compare reconstructed object trajectories against a 200 FPS reference camera; if the staggered array does not reduce trajectory or rendering error relative to the synchronized array, the effective-frame-rate claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that high-speed 4D reconstruction does not require high-speed cameras: staggering the capture start times of commodity cameras multiplies the system's temporal sampling rate by the group size, and a video-diffusion-based artifact-fix model converts the resulting sparse-view reconstructions into clean, temporally consistent renderings. With twelve 25 FPS cameras divided into four groups the system effectively samples at 100 FPS, and with eight groups at 200 FPS, then refines the 4D Gaussian model with the diffusion output. On DNA-Rendering and Neural3DV, where videos are temporally subsampled to simulate large inter-frame motion, the full method beats synchronized GS4D, 4DGS, and K-Planes on PSNR, SSIM, and LPIPS, and it also outperforms RIFE and MoMo frame-interpolation pipelines. The authors further release a dataset of 12 asynchronously captured sequences of dancing, sports, and fast object motion.

Load-bearing premise

The load-bearing premise is that temporally subsampling existing 4D datasets faithfully mimics real asynchronous capture, so the gains measured on synthetic subsampling will transfer to physical trigger-delay capture.

Editorial extensions

If this is right

  • Existing 15-30 FPS camera arrays can be upgraded to capture fast-dynamic scenes at 4-8x temporal resolution by adding trigger-delay control, with no additional camera hardware.
  • Asynchronous multi-view sampling, rather than frame interpolation, becomes a practical route to reconstructing large, nonlinear motion such as cloth movement and sports actions.
  • Video-diffusion refinement can serve as a general post-processor for sparse-view dynamic reconstructions, fixing floaters while preserving temporal consistency across frames.
  • The 12-sequence asynchronous dataset provides a first benchmark for evaluating 4D reconstruction under true staggered-capture conditions.
  • Dividing the camera array into four or eight groups yields effective capture rates of 100 or 200 FPS respectively, per the timing formula.

Reading between the lines

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

  • Inference: If the same artifact-fix model were trained on non-human motion rather than the human-centric DNA-Rendering data, the async-capture scheme would likely generalize to animal, vehicle, and robot scenes; the paper explicitly leaves this unverified.
  • Inference: Per-scene fine-tuning of the artifact-fix model, which the paper reports improves fine dress details but adds time overhead, could be combined with a cheaper leave-one-out schedule to make it practical as a default.
  • Inference: Because async capture trades viewpoint density for temporal density, one might pair it with multi-view depth or flow priors to cut the diffusion model's hallucination risk in highly textured regions.
  • Inference: An immediate stress test is to apply the same staggering scheme with fewer groups or irregular delays to find the minimum temporal sampling needed for given motion speeds.
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 / 5 minor

Summary. The paper proposes a 4D reconstruction system for fast-dynamic scenes that combines an asynchronous multi-camera capture scheme with a video-diffusion-based artifact-fix model. On the capture side, cameras are divided into groups whose start times are staggered, so that a 25 FPS camera array can provide temporal samples equivalent to 100-200 FPS. On the reconstruction side, an initial 4D Gaussian Splatting (GS4D) model is trained on the asynchronous views, rendered videos are refined by a LoRA-fine-tuned Wan2.1 video diffusion model, and the refined videos supervise a second stage of GS4D optimization. Experiments compare against K-Planes, 4DGS, and GS4D on temporally subsampled DNA-Rendering and Neural3DV datasets, with ablations on asynchronous versus synchronous capture and on the artifact-fix model, plus a qualitative evaluation on a new real 12-camera asynchronous capture dataset.

Significance. If validated, the asynchronous capture scheme is a genuinely low-cost way to increase temporal sampling for 4D reconstruction, and the first real asynchronous multi-view 4D dataset is a useful community resource. The video-diffusion artifact-fix model directly addresses the temporal inconsistency that image-diffusion refinement would introduce, and the ablations support the qualitative direction that asynchronous capture plus artifact fixing improves reconstruction over synchronous capture. The paper's central quantitative claims, however, rest on a synthetic subsampling protocol and on evaluations that overlap with the diffusion model's training data; the real-capture evidence is qualitative only. These issues are fixable, but they currently prevent the 100-200 FPS hardware claim from being fully established.

major comments (4)
  1. [Sec. 4.1 and Sec. 4.2] The artifact-fix video diffusion model is trained on 750 noisy-clean pairs constructed from the DNA-Rendering dataset (Sec. 4.1), and the quantitative comparison in Table 1 and the ablation in Table 3 are also on DNA-Rendering. Because the test frames come from the same dataset distribution as the training pairs, the reported gains on DNA-Rendering do not by themselves establish that the diffusion model generalizes; they could partly reflect memorization of the training distribution. The authors should report results on a subject- or sequence-disjoint split of DNA-Rendering, or provide additional cross-dataset quantitative evidence beyond Table 2, to support the claim that the model repairs artifacts rather than memorizes training scenes.
  2. [Sec. 4.2 and Fig. 7] The real 12-camera asynchronous dataset is evaluated only qualitatively. No quantitative metrics (PSNR, SSIM, LPIPS, or temporal accuracy) are reported on the real async sequences, and the comparison with synchronous capture uses separately recorded sequences with similar motion rather than the same motion. The central hardware claim that commodity 25 FPS cameras can achieve 100-200 FPS equivalent temporal resolution is therefore not quantitatively validated on real asynchronous captures. Please report metrics on the real async dataset, for example by holding out one camera view and computing novel-view synthesis quality, and specify the exposure time, trigger delay accuracy, and global-versus-rolling shutter mode of the capture system.
  3. [Sec. 4.2] The simulation protocol treats temporally subsampled frames as instantaneous samples at their declared timestamps, but real asynchronous capture at 25 FPS with K=4 or K=8 implies stagger intervals of 10 ms or 5 ms, and the claimed frame-rate equivalence is only meaningful if exposure is much shorter than the stagger interval and if trigger timing is accurate to a comparable degree. The paper does not report exposure settings or trigger precision for the 12-camera array, and the synthetic subsampled frames do not contain the motion blur or rolling-shutter distortion that real asynchronous captures would exhibit. A validation experiment comparing the artifact-fix model on real async frames against the corresponding GS4D renders, or a sensitivity analysis with respect to trigger jitter and exposure duration, would substantiate the claim that the synthetic subsampling is a faithful proxy for real async capture.
  4. [Sec. 4.3 and Tables 1-3] Tables 1-3 report a single value for each method and metric, with no error bars or number of repeated runs. The differences between the proposed method and the strongest baseline (e.g., PSNR 26.76 vs. 24.75 on DNA-Rendering) could be within run-to-run optimization variability. Please report means and standard deviations over multiple seeds or runs, or otherwise justify that the reported differences are statistically meaningful.
minor comments (5)
  1. [Abstract and Sec. 3.2] The phrase 'equivalent frame rate of 100-200 FPS' should be clarified: the staggering increases the temporal sampling density per camera group, but the total number of image samples per second is unchanged and the number of views at each timestamp is reduced by a factor of K. The paper does note the sparse-view issue, but the abstract's wording is likely to mislead readers.
  2. [Sec. 3.4] The sentence 'By applying our per-scene artifact-fix diffusion model to refine the Gaussian representation' conflicts with Sec. 5.3, where per-scene fine-tuning is described as an optional variant that is not adopted by default. Please reword to distinguish the default model from the per-scene fine-tuned variant.
  3. [Sec. 2.1] The citation '[jiang2024timeformer]' for TimeFormer has no corresponding entry in the reference list; please add the reference or remove the citation.
  4. [Sec. 3.3] There is a typo in 'a artifact-fix video diffusion model' (should be 'an artifact-fix video diffusion model'), and 'Wan-VAEE' appears to be a typo for 'Wan-VAE'.
  5. [Sec. 5.3 and Fig. 11] In Fig. 11, the panel label 'w/oper-scene' should be 'w/o per-scene'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the asynchronous-capture gains are empirical comparisons against external baselines and ground truth, while the frame-rate equivalence is definitional and the diffusion self-distillation loop is not load-bearing.

full rationale

The paper's central result is an empirical system: asynchronous capture is defined by staggering camera start times, and the claimed benefit is tested by comparing asynchronous versus synchronous inputs on DNA-Rendering, Neural3DV, and a real 12-camera capture. The 'equivalent frame rate of 100–200 FPS' statement (Sec. 3.2, t_i = i·(τ/K)+j·τ) is a definitional consequence of temporal interleaving, not a derived empirical prediction, and no downstream quantitative claim is obtained solely from that arithmetic. The artifact-fix diffusion model is trained on GS4D-rendered noisy/clean pairs synthesized by temporal subsampling (Sec. 3.3), and Eq. 4 later uses the model's refined videos as supervision for the Gaussian optimization; this is a mild self-distillation loop. However, the reported improvements do not reduce to that loop by construction: Tables 1–2 report PSNR/SSIM/LPIPS against ground-truth frames from a held-out view, and Neural3DV is not part of the diffusion model's training set. The baselines (K-Planes, 4DGS, GS4D, RIFE, MoMo) are external methods, and the async-versus-sync ablation in Table 3 is an empirical result of providing staggered timestamps to a 4D representation, not a tautology. No load-bearing step is justified by self-citation, no uniqueness theorem is imported from the authors, and no result is equivalent to its input by definition. The mild self-training loop and the synthetic-subsampling proxy for real asynchronous capture are validity concerns, but they are not circularity in the sense of the analysis.

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

The claims rest on the transfer of simulated subsampling to real capture and on a diffusion self-distillation loop. Hyperparameters such as LoRA rank, learning rate, iteration schedule, and subsampling factors are chosen by hand, and the simulation transfer is an explicit modeling assumption stated in Sec 4.2. There are no invented physical entities.

free parameters (6)
  • LoRA rank = 16
    Chosen for the artifact-fix diffusion fine-tune (Sec 4.1); sets model capacity and is not derived from data.
  • Diffusion fine-tuning learning rate = 1e-4
    Reported in Sec 4.1 as a hand-set hyperparameter for Wan2.1 LoRA training.
  • GS4D optimization schedule = 7k iterations before, 7k with diffusion supervision
    Sec 4.1 defines the split between initial reconstruction and refinement; the checkpoint choice affects artifact severity and final metrics.
  • DNA-Rendering temporal subsampling factor = 4x
    Sec 4.2 uses 4x subsampling to 'induce large motions'; the difficulty of the synthetic benchmark is controlled by this choice.
  • Neural3DV temporal subsampling factor = 12x
    Sec 4.2 uses 12x subsampling for Neural3DV; no sensitivity analysis is provided.
  • Number of camera groups K = 4 (default), 8 for 200 FPS
    Sec 3.2 sets K=4 in experiments, trading temporal density against per-timestamp viewpoint count.
assumptions (6)
  • domain assumption GS4D 4D Gaussian splatting with 4D spherical harmonics is a sound base representation for reconstructing the dynamic scenes.
    The paper builds directly on GS4D (Sec 3.1, Eq. 1-2) and does not re-derive or validate the representation against other 4D bases.
  • ad hoc to paper Temporal subsampling of existing datasets faithfully simulates both large motion and the asynchronous capture pattern.
    Sec 4.2 states that temporal downsampling 'effectively simulate[s] large inter-frame motion'; the simulation is the main evidence for the capture claim.
  • ad hoc to paper Artifacts produced by GS4D trained on subsampled data are representative of artifacts from real asynchronous capture.
    The diffusion training pairs in Sec 4.1 are generated from subsampled DNA-Rendering, then applied to real async data in Fig 7 without a quantitative check of artifact distribution.
  • domain assumption Wan2.1 video diffusion priors transfer to 4D reconstruction renderings and can be fine-tuned with LoRA without catastrophic forgetting.
    Sec 3.3 and 3.4 rely on the pretrained video model generalizing to rendered, artifact-corrupted videos; no probing of the prior's limits beyond a limited qualitative comparison.
  • domain assumption Using diffusion-sampled videos as pseudo-ground-truth in Eq. 4 improves the Gaussian model without systematically biasing geometry.
    The refinement loop treats the diffusion output as clean supervision; the paper does not measure geometric drift or hallucination on real scenes.
  • domain assumption The custom camera array applies trigger delays accurately enough to realize the staggered timing described in Sec 3.2.
    Sec 4.2 says delays are 'manually introduced' but provides no temporal calibration or synchronization error measurement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 4DSloMo: 4D Reconstruction for High Speed Scene with Asynchronous Capture." pith.science (2026). https://pith.science/paper/HM5WQZ7L

@misc{pith2026250705163,
  author       = {Pith},
  title        = {Pith review of: 4DSloMo: 4D Reconstruction for High Speed Scene with Asynchronous Capture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HM5WQZ7L}},
  note         = {Machine review of arXiv:2507.05163}
}
read the original abstract

Reconstructing fast-dynamic scenes from multi-view videos is crucial for high-speed motion analysis and realistic 4D reconstruction. However, the majority of 4D capture systems are limited to frame rates below 30 FPS (frames per second), and a direct 4D reconstruction of high-speed motion from low FPS input may lead to undesirable results. In this work, we propose a high-speed 4D capturing system only using low FPS cameras, through novel capturing and processing modules. On the capturing side, we propose an asynchronous capture scheme that increases the effective frame rate by staggering the start times of cameras. By grouping cameras and leveraging a base frame rate of 25 FPS, our method achieves an equivalent frame rate of 100-200 FPS without requiring specialized high-speed cameras. On processing side, we also propose a novel generative model to fix artifacts caused by 4D sparse-view reconstruction, as asynchrony reduces the number of viewpoints at each timestamp. Specifically, we propose to train a video-diffusion-based artifact-fix model for sparse 4D reconstruction, which refines missing details, maintains temporal consistency, and improves overall reconstruction quality. Experimental results demonstrate that our method significantly enhances high-speed 4D reconstruction compared to synchronous capture.

Figures

Figures reproduced from arXiv: 2507.05163 by the authors.

Figure 1
Figure 1. Our 4D Reconstruction Results of the real-captured scene. We propose an asynchronous capture scheme, which increases the effective capture frame rate by staggering the start times of cameras without any additional cost. We further leverage video diffusion priors to enhance the reconstruction results. The results show that our method can reconstruct high speed and complex motion with high quality. Reconstructing fast… view at source ↗
Figure 2
Figure 2. The overall pipeline of our model. Given several asynchronous multi-view videos, we first initialize a 4D Gaussian model for a specific iteration. We then employ an artifact-fix video diffusion model to refine the input videos. The refined videos are subsequently used to update the 4D Gaussian model. stereo [Vlasic et al. 2009] offer higher quality, they typically require complex lighting and are costly. To improve … view at source ↗
Figure 3
Figure 3. Illustration of artifact-fix video diffusion model setup. Our model freezes all parameters in the network, except for the LoRA weights, to fine-tune a video diffusion model. Precisely, we integrate Lora parameters into the DiT model. With a LoRA rank designated as 16, this integration takes place in each transformer block. starting at different times for the capture: the 𝑖-th camera starts capturing an image at the … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Qualitative result on the Neural3DV dataset [Li et al. 2022]. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative result on the DNA-Rendering dataset [Cheng et al. 2023a]. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the capture system. consists of all frames from a held-out view. Neural3DV captures multi-view videos at 30 FPS, each lasting ten seconds. We subsample the videos to one-twelfth of their original frame rate. For each scene, one view is held out for test…
Figure 7
Figure 7. Figure 7: Quality result on our real-capture dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Ablation study of asynchronous capture and diffusion model on the DNA-Rendering dataset [Cheng et al [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 7
Figure 7. Figure 7: Since we are unable to capture both synchronous and asyn [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Ablation study of asynchronous capture and diffusion model on LongVolCap datasets [Xu et al [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Ablation study on video diffusion model and image diffusion model. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 28 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering

    Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982--19993, 2023 a

    work page 2023

  3. [3]

    Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering

    Wei Cheng, Ruixiang Chen, Wanqi Yin, Siming Fan, Keyu Chen, Honglin He, Huiwen Luo, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. arXiv pre...

    arXiv 2023

  4. [4]

    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, 2022

    arXiv 2022

  5. [5]

    Cat3d: Create anything in 3d with multi-view diffusion models

    Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024

    arXiv 2024

  6. [6]

    Timeformer: Capturing temporal relationships of deformable 3d gaussians for robust reconstruction

    DaDong Jiang, Zhihui Ke, Xiaobo Zhou, Zhi Hou, Xianghui Yang, Wenbo Hu, Tie Qiu, and Chunchao Guo. Timeformer: Capturing temporal relationships of deformable 3d gaussians for robust reconstruction. arXiv preprint arXiv:2411.11941, 2024

    work page arXiv 2024

  7. [7]

    A compact dynamic 3d gaussian representation for real-time dynamic view synthesis

    Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis. In European Conference on Computer Vision, pages 394--412. Springer, 2024

    work page 2024

  8. [8]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk \"u hler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42 0 (4): 0 139--1, 2023

    2023

Show all 40 references
  1. [9]

    Neural 3d video synthesis from multi-view video

    Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF conference on computer vision...

    2022

  2. [10]

    Spacetime gaussian feature splatting for real-time dynamic view synthesis

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508--8520, 2024

    2024

  3. [11]

    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. In SIGGRAPH Asia 2022 Conference Papers, pages 1--9, 2022

    2022

  4. [12]

    Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle

    Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136--21145, 2024

    2024

  5. [13]

    Efficient space-time sampling with pixel-wise coded exposure for high-speed imaging

    Dengyu Liu, Jinwei Gu, Yasunobu Hitomi, Mohit Gupta, Tomoo Mitsunaga, and Shree K Nayar. Efficient space-time sampling with pixel-wise coded exposure for high-speed imaging. IEEE transactions on pattern analysis and machine intelligence, 36 0 (2): 0 248--260, 2013

    2013

  6. [14]

    Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models

    Xinhang Liu, Jiaben Chen, Shiu-hong Kao, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. 2023

    2023

  7. [15]

    3d geometry-aware deformable gaussian splatting for dynamic view synthesis

    Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900--8910, 2024

    2024

  8. [16]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65 0 (1): 0 99--106, 2021

    2021

  9. [17]

    Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs

    Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pag...

    2022

  10. [18]

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

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

    2022 arXiv

  11. [19]

    D-nerf: Neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318--10327, 2021

    2021

  12. [20]

    K-planes: Explicit radiance fields in space, time, and appearance

    Sara Fridovich-Keil and Giacomo Meanti , Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023

    2023

  13. [21]

    D \"a rf: Boosting radiance fields from sparse input views with monocular depth adaptation

    Jiuhn Song, Seonghoon Park, Honggyu An, Seokju Cho, Min-Seop Kwak, Sungjin Cho, and Seungryong Kim. D \"a rf: Boosting radiance fields from sparse input views with monocular depth adaptation. Advances in Neural Information Processing Systems, 36: 0 68458--68470, 2023

    2023

  14. [22]

    Sparf: Neural radiance fields from sparse and noisy poses

    Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 2023

    2023

  15. [23]

    Wan: Open and advanced large-scale video generative models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pande...

    2025 arXiv

  16. [24]

    Shape of motion: 4d reconstruction from a single video

    Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024

    2024

  17. [25]

    Vidu4d: Single generated video to high-fidelity 4d reconstruction with dynamic gaussian surfels

    Yikai Wang, Xinzhou Wang, Zilong Chen, Zhengyi Wang, Fuchun Sun, and Jun Zhu. Vidu4d: Single generated video to high-fidelity 4d reconstruction with dynamic gaussian surfels. Advances in Neural Information Processing Systems, 37: 0 131316--131343, 2025

    2025

  18. [26]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310...

    2024

  19. [27]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310--20320, 2024 b

    2024

  20. [28]

    Difix3d+: Improving 3d reconstructions with single-step diffusion models

    Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3d+: Improving 3d reconstructions with single-step diffusion models. arXiv preprint arXiv: 2503.01774, 2025

    2025 arXiv

  21. [29]

    Reconfusion: 3d reconstruction with diffusion priors

    Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF conference on computer vision and patt...

    2024

  22. [30]

    Diffusionerf: Regularizing neural radiance fields with denoising diffusion models

    Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180--4189, 2023

    2023

  23. [31]

    4k4d: Real-time 4d view synthesis at 4k resolution

    Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In CVPR, 2024 a

    2024

  24. [32]

    4k4d: Real-time 4d view synthesis at 4k resolution

    Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20029--20040, 2024 b

    2024

  25. [33]

    Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting

    Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv e-prints, pages arXiv--2402, 2024 a

    2024

  26. [34]

    Freenerf: Improving few-shot neural rendering with free frequency regularization

    Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254--8263, 2023

    2023

  27. [35]

    Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331--20341, 2024 b

    2024

  28. [36]

    Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting

    Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024 c

    2024

  29. [37]

    Lm-gaussian: Boost sparse-view 3d gaussian splatting with large model priors

    Hanyang Yu, Xiaoxiao Long, and Ping Tan. Lm-gaussian: Boost sparse-view 3d gaussian splatting with large model priors. arXiv preprint arXiv:2409.03456, 2024

    2024 arXiv

  30. [38]

    Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction

    Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35: 0 25018--25032, 2022

    2022

  31. [39]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836--3847, 2023

    2023

  32. [40]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586--595, 2018

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.