Pith. sign in

REVIEW 3 major objections 5 minor 44 references

A single in-the-wild smoke video can be reconstructed into a dynamic, editable 3D asset with a +2.22 dB PSNR advantage over prior methods.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 17:05 UTC pith:2TM5OIR5

load-bearing objection Genuine pipeline contribution for in-the-wild smoke extraction, but the headline +2.22 dB gain is not direct evidence of 3D correctness—the one controlled experiment shows the multi-view decoupling hurts the exact metric used for wild videos. the 3 major comments →

arxiv 2509.11114 v1 pith:2TM5OIR5 submitted 2025-09-14 cs.CV cs.LG

WildSmoke: Ready-to-Use Dynamic 3D Smoke Assets from a Single Video in the Wild

Yuqiu Liu , Jialin Song , Manolis Savva , Wuyang Chen This is my paper
classification cs.CV cs.LG
keywords smoke reconstructiondynamic 3D assetsin-the-wild videomulti-view generationGaussian splattingfluid simulationdehazingnovel view synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes a pipeline that takes a single video of smoke captured with a moving camera in an unconstrained setting—cluttered background, unknown camera pose, single viewpoint trajectory—and reconstructs a dynamic 3D smoke field that can be re-rendered from new views, extrapolated into the future, and seeded into a fluid simulator for editing. It identifies three obstacles: background noise visible through translucent smoke, unknown camera poses, and the coupling between camera viewpoint and time in a single video. The pipeline combines smoke segmentation with dehazing, monocular geometry-based initialization of particle positions and camera poses, and generated multi-view supervision plus local pose perturbation to decouple space and time. The authors report an average +2.22 dB PSNR improvement over previous reconstruction and generation methods on real-world smoke videos (FLAME and Pixabay), and demonstrate interactive simulation with wind and obstacles. If correct, the method turns ordinary amateur or drone footage of smoke into ready-to-use 4D assets for VFX.

Core claim

On the paper's own terms, the central claim is that faithful, dynamic 3D smoke reconstruction from a single in-the-wild video is possible, provided that background contamination is removed, geometry and pose are initialized from a coarse monocular 3D estimate, and the one-to-one coupling between camera viewpoint and timestep is broken by generative multi-view frames and local pose perturbation. Using Gaussian-splatting particles for density and position-based fluid simulation for physical particles, the reconstructed asset supports both novel-view synthesis and future-frame prediction, and can be edited by simulating external forces and obstacles. The paper demonstrates this on synthetic gro

What carries the argument

The pipeline's load-bearing elements are: (1) smoke extraction that combines one-shot instance segmentation (SAM annotation propagated by SegGPT) with a dark-channel dehazing step to separate semi-transparent light smoke from visible backgrounds; (2) initialization from a pretrained monocular geometry estimator (DUSt3R) that yields both per-frame camera poses and a sparse foreground point cloud from which physical and visual particles are seeded; (3) multi-view supervision generated by SV4D 2.0 at fixed azimuth offsets, with exponentially decaying weights over frame index to suppress structural collapse in later generated frames; and (4) local pose perturbation, which associates the pose at

Load-bearing premise

The decoupling between viewpoint and time is only as trustworthy as the synthetic multi-view frames: the generated videos and the model's own perturbed renders are treated as valid training targets, yet no ground-truth novel views exist to confirm that these targets carry true 3D structure rather than the original single-view bias.

What would settle it

Capture a smoke plume with two synchronized cameras whose poses are calibrated, train with one camera's video as the single in-the-wild input (hiding the other), and compare the pipeline's rendered frames from the second camera's viewpoint against that camera's actual footage at corresponding timesteps; if the rendered views differ by more than the reported PSNR gap, or if removing SV4D-generated frames produces no degradation in future prediction, the central claim is wrong.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reconstruction is faithful, a single handheld or drone video of smoke yields a complete 4D asset: density field, velocity field, and particle geometry, without multi-camera rigs or lab backgrounds.
  • Novel-view and future-frame rendering both improve over baselines, meaning the recovered velocity field is not just a per-frame fit but captures enough temporal dynamics to advect particles into unseen frames.
  • The same asset can be loaded into a standard incompressible-flow simulator, enabling content editing such as global/local wind and rigid obstacles, a direct route from real footage to VFX.
  • The +2.22 dB average PSNR gain on in-the-wild videos is a concrete quantitative target for future methods to beat.
  • The smoke-extraction and dehazing steps are reusable components that can be dropped into other reconstruction pipelines.

Where Pith is reading between the lines

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

  • The method's reliance on a generative model (SV4D 2.0) for multi-view supervision means its novel-view accuracy in the wild is only as good as the generator's cross-view consistency; the paper's own appendix shows structural collapse at later timesteps, so the exponentially decaying weights are an admission that the strongest supervision arrives early. A more temporally stable generative prior cou
  • If the synthetic-supervision premise is correct, a similar extraction-plus-decoupling recipe might generalize to other semi-transparent volumetric phenomena (fire, haze, spray, steam) where background and foreground are hard to separate.
  • A testable extension: apply the pipeline to a video where an independent second camera provides ground-truth novel views; the gap between generated and actual novel views would quantify how much of the PSNR gain is genuine 3D structure versus input-view consistency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes WildSmoke, a full pipeline for turning a single in-the-wild video of smoke into an editable 4D smoke asset. The pipeline consists of: smoke segmentation via SAM plus SegGPT, dehazing for light smoke, DUSt3R-based pose and particle initialization, SV4D 2.0-based multi-view generation with exponentially decaying weights, Gaussian-particle training with PBF regularization, and local pose perturbation to decouple viewpoint from time. The authors evaluate on synthetic data with ground-truth novel views and on FLAME plus three Pixabay videos, reporting a +2.22 dB average PSNR improvement over FluidNexus and HyFluid on the wild input-view future-prediction task, and they demonstrate simulation-based editing of the reconstructed smoke. The central claim is that this pipeline faithfully reconstructs accurate, dynamic, view-consistent 3D smoke from a single monocular in-the-wild video.

Significance. If the central claim holds, the paper makes a useful practical contribution: it extends fluid reconstruction beyond laboratory multi-view settings to casual single-camera footage and produces assets that can be re-simulated. The paper has genuine strengths: the ablation in Table 2 is cumulative and mostly consistent; Table 6 in the supplement provides a clean demonstration that the proposed smoke extraction dramatically rescues the baselines (e.g., FluidNexus rises from 6.67 to 21.78 dB on FLAME); and the synthetic experiments offer a controlled check with ground-truth novel views and future frames. The pipeline is also reasonably economical in compute (Table 5). However, the evidence for faithful 3D structure in the wild is indirect: the only ground-truth validation is on synthetic data, while all wild-video numbers use input-view future prediction. The paper does not fully close the gap between these two regimes, and one internal ablation result directly weakens the link.

major comments (3)
  1. [Table 2 and Sec. 4.3] There is a striking non-monotonicity in the cumulative ablation. Adding '+ Multi-Views' to '+ Local Perturbation' improves future prediction at the novel view (23.59 to 25.04 dB) but reduces future prediction at the input view by 1.59 dB (26.85 to 25.26). Since Sec. 4.3 states that all wild-video comparisons (FLAME and Pixabay) use only future prediction at the input view, the very module intended to decouple viewpoint from time is shown, in the controlled experiment, to hurt the metric used to support the +2.22 dB headline claim. The text says the components 'progressively improve PSNR' but never mentions this regression. Please either justify why the synthetic input-view regression should not transfer to the wild setting, or report wild results for the configuration without multi-view supervision; otherwise the wild gain cannot be attributed to the multi-view decoupling component.
  2. [Sec. 3.4 and Appendix A.5/B.2] The 'local pose perturbation' step uses the model's own rendering as the target for the perturbed pose (R(t+Δ_t) mod T, t). This is a self-supervision loop: it enforces internal consistency but provides no independent evidence about the true 3D scene. Moreover, the only external multi-view source, SV4D 2.0, is acknowledged in Appendix B.2 and Fig. 12 to structurally collapse at later timesteps, and Appendix A.5 therefore down-weights those frames exponentially. Thus the model is trained to agree with itself at nearby poses and with increasingly noisy generated views. The paper needs an independent validation of the 3D structure on real footage—for example, held-out real multi-view frames from a known camera motion, or a geometric consistency check such as measuring reprojection error against independently estimated depth—before the claims 'faithfully reconstruct accurate and dynamic 3D s
  3. [Tables 3, 4 and Table 6] The +2.22 dB average improvement is computed from single-run PSNR values per video (gains of 1.10, 0.50, 3.98, and 3.28 dB on four videos). No error bars, repeated runs, or statistical significance are reported. In addition, Table 6 shows that simply applying the proposed smoke extraction to the baselines boosts FluidNexus from 6.67 to 21.78 dB on FLAME and HyFluid from 10.37 to 21.67 dB. This suggests that a substantial portion of the improvement over prior methods may come from preprocessing rather than from the new reconstruction or decoupling techniques. The paper should report variance and, more importantly, compare all methods under identical foreground extraction so the marginal contribution of the reconstruction pipeline is isolated.
minor comments (5)
  1. [Sec. 4.1] The synthetic dataset is described as containing 'two' rendered videos, but Table 1 reports only aggregate PSNR. Please state how many synthetic scenes/videos are averaged and provide per-scene numbers or confidence intervals.
  2. [Table 5] The 'Total' row is difficult to read: '3.44 5.84 4.36' appears to have inconsistent column alignment. Also, the per-stage GPU hours are not separated into training/inference cleanly for columns after the first.
  3. [Appendix A.6] The learnable buoyancy coefficient is a potentially important free parameter and appears only in the supplement. It should be described in the main methods section, and its effect on the reported results should be ablated or at least explicitly listed as a tuned parameter.
  4. [Sec. 3.3 and Appendix A.5] The exponentially decaying weight for generated frames (w_min = 0.0, k = 0.02) is introduced in the supplement but not discussed in the main text. Readers need to know that early generated frames dominate and later ones are nearly ignored, since this bears on the reliability of the multi-view supervision.
  5. [Sec. 4.4] The simulation demos are qualitative. A quantitative evaluation, even a simple metric comparing simulated density evolution to the held-out future frames, would strengthen the claim that the reconstructed assets are physically plausible.

Circularity Check

0 steps flagged

No significant circularity: the headline PSNR gains are measured on held-out future frames and independent synthetic ground truth; the self-supervised local-pose-perturbation target is a regularizer, not a fitted prediction.

full rationale

The paper's derivation chain is not circular. The central quantitative claim (+2.22 dB on in-the-wild videos, Sec. 4.3) is computed on future frames t=241..270 that are never used in training (Sec. 4.1: 'No model is ever trained with ground-truth future frames from videos'), and the only metric available on wild videos is future prediction at the input view. On synthetic videos, both novel-view synthesis and future prediction are evaluated against ground-truth renders from a second camera trajectory (Sec. 4.1). These evaluations are independent of the model's training targets. The local pose perturbation (Sec. 3.4) uses the model's own rendering as the target for perturbed viewpoints ('its corresponding rendering result as the target'), which is self-supervision rather than external evidence; however, it is a training-time regularizer that enforces self-consistency, not a prediction that reduces to a fit. The SV4D-generated multi-view frames are explicitly recognized as unreliable at later timesteps (Appendix B.2) and are exponentially down-weighted (Appendix A.5), so the pipeline does not treat its own generations as ground truth. There is no load-bearing self-citation: the compared baselines (HyFluid, FluidNexus) are external works, and the authors' own contributions are presented with ablations against ground truth. The reported wild-video gain could be influenced by the smoke-extraction pre-processing (Table 6 shows baselines improve substantially with it), but that is an attribution/validity concern, not circularity. No equation in the paper defines a predicted quantity in terms of a fitted parameter by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claim rests on roughly six hand-chosen or fitted quantities, of which the learnable buoyancy (A.6) is the most consequential because it couples the physical extrapolation to the training data. The axioms are all domain assumptions inherited from graphics and vision priors; none is introduced ad hoc to the paper except possibly the assumption that SV4D generations are reliable supervision, which the paper itself qualifies. No new physical entities are postulated.

free parameters (6)
  • Learnable buoyancy coefficient = not reported
    Appendix A.6 makes the buoyancy strength in the PBF simulation learnable and optimizes it jointly with reconstruction; it directly shapes the velocity field used for future prediction, the main evaluation metric.
  • SV4D novel-view angle set = azimuth offsets -10, 10, 20, 30 degrees
    Section 3.3: hand-chosen angles for generative multi-view synthesis; the choices determine the spatial extent of synthetic supervision and are not justified by scene coverage.
  • Exponential decay parameters for generated frames = k = 0.02, w_min = 0.0, t_0 = 0
    Appendix A.5: hand-chosen schedule for down-weighting SV4D-generated frames due to their structural collapse over time.
  • Local pose perturbation schedule = Delta t increased from 2 to 4
    Section 3.4: hand-chosen perturbation range, progressively increased during training.
  • Frequency loss weight = lambda_freq = 0.001 with linear warm-up
    Appendix A.7: hand-chosen weight for the FFT amplitude and phase loss.
  • Particle count after voxel downsampling = 100 to 300 points per video
    Appendix A.3: voxel grid downsampling controls the initial particle count; this capacity choice interacts with reconstruction quality.
axioms (6)
  • domain assumption Atmospheric scattering model I = I_clean * T + A * S (Eq. 1), with the same model used to synthesize dehazing training data and to invert the smoke layer (Eq. 2).
    Section 3.1: the clean background recovered by the fine-tuned dehazer is divided out; if the blend model does not match real smoke translucency, the extracted foreground S inherits systematic error.
  • domain assumption Dark channel prior estimates atmospheric light A.
    Section 3.1: inherited from He et al. [13]; accuracy on smoke rather than haze is not validated.
  • domain assumption SV4D 2.0 generated multi-view frames are valid supervision for near-baseline viewpoints.
    Section 3.3 and Appendix A.5: the paper relies on generated views while also showing they structurally collapse at later timesteps (Appendix B.2), hence exponential down-weighting; reliability is assumed in the regime where it matters.
  • domain assumption Smoke is an incompressible fluid adequately regularized by position-based fluid simulation with an incompressibility constraint.
    Section 3.4: the physical particle field is regularized by PBF; this graphics approximation is not validated against real turbulent smoke.
  • domain assumption DUSt3R monocular geometry provides camera poses and a foreground point cloud accurate enough to initialize particles.
    Section 3.2 and Appendix A.2: the paper states that without this initialization training may fail to converge; errors here propagate into the recovered fields.
  • standard math Axis convention conversion between DUSt3R and Gaussian Splatting is a fixed diagonal flip diag(1,-1,-1) (and diag(1,-1,-1,1) for poses).
    Appendix A.2 and A.3: a coordinate convention rather than an empirical claim.

pith-pipeline@v1.3.0-alltime-deepseek · 14121 in / 22504 out tokens · 242819 ms · 2026-08-04T17:05:23.202494+00:00 · methodology

0 comments
read the original abstract

We propose a pipeline to extract and reconstruct dynamic 3D smoke assets from a single in-the-wild video, and further integrate interactive simulation for smoke design and editing. Recent developments in 3D vision have significantly improved reconstructing and rendering fluid dynamics, supporting realistic and temporally consistent view synthesis. However, current fluid reconstructions rely heavily on carefully controlled clean lab environments, whereas real-world videos captured in the wild are largely underexplored. We pinpoint three key challenges of reconstructing smoke in real-world videos and design targeted techniques, including smoke extraction with background removal, initialization of smoke particles and camera poses, and inferring multi-view videos. Our method not only outperforms previous reconstruction and generation methods with high-quality smoke reconstructions (+2.22 average PSNR on wild videos), but also enables diverse and realistic editing of fluid dynamics by simulating our smoke assets. We provide our models, data, and 4D smoke assets at [https://autumnyq.github.io/WildSmoke](https://autumnyq.github.io/WildSmoke).

Figures

Figures reproduced from arXiv: 2509.11114 by Jialin Song, Manolis Savva, Wuyang Chen, Yuqiu Liu.

Figure 1
Figure 1. Figure 1: Challenges of smoke reconstruction from a single in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: To reconstruct smoke from a single video in the wild, our pipeline includes five steps: 1) Smoke extraction, with background [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Smoke extraction, with background removal for light smoke. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Local pose perturbation. Given the original camera pose [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of novel view synthesis and future predictions on [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of future predictions (input view) on the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of future predictions (input view) on videos [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualizations of simulations with our smoke assets in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Camera poses and foreground smoke points (for the [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualizations of extracted smoke from wild videos. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Novel view synthesis on wild videos. Our reconstruc [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Ablation study of novel view synthesis and future predictions on synthetic smoke videos. [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 9 linked inside Pith

  1. [1]

    Dy- namic upsampling of smoke through dictionary-based learn- ing.ACM Transactions on Graphics (TOG), 40(1):1–19, 2020

    Kai Bai, Wei Li, Mathieu Desbrun, and Xiaopei Liu. Dy- namic upsampling of smoke through dictionary-based learn- ing.ACM Transactions on Graphics (TOG), 40(1):1–19, 2020

    2020

  2. [2]

    Enginebench: flow reconstruction in the transparent combustion chamber iii optical engine.arXiv preprint arXiv:2406.03325, 2024

    Samuel J Baker, Michael A Hobley, Isabel Scherl, Xiaohang Fang, Felix CP Leach, and Martin H Davy. Enginebench: flow reconstruction in the transparent combustion chamber iii optical engine.arXiv preprint arXiv:2406.03325, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Cambridge University Press, 2006

    Andrew Bennett.Lagrangian fluid dynamics. Cambridge University Press, 2006

    2006

  4. [4]

    Learning meaningful controls for fluids.ACM Transactions on Graphics (TOG), 40(4):1–13, 2021

    Mengyu Chu, Nils Thuerey, Hans-Peter Seidel, Christian Theobalt, and Rhaleb Zayer. Learning meaningful controls for fluids.ACM Transactions on Graphics (TOG), 40(4):1–13, 2021

    2021

  5. [5]

    Physics informed neural fields for smoke reconstruction with sparse data.ACM Transactions on Graphics (ToG), 41(4):1–14, 2022

    Mengyu Chu, Lingjie Liu, Quan Zheng, Erik Franz, Hans- Peter Seidel, Christian Theobalt, and Rhaleb Zayer. Physics informed neural fields for smoke reconstruction with sparse data.ACM Transactions on Graphics (ToG), 41(4):1–14, 2022

    2022

  6. [6]

    Learn- ing vortex dynamics for fluid inference and prediction.arXiv preprint arXiv:2301.11494, 2023

    Yitong Deng, Hong-Xing Yu, Jiajun Wu, and Bo Zhu. Learn- ing vortex dynamics for fluid inference and prediction.arXiv preprint arXiv:2301.11494, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Fluid simulation on neural flow maps.ACM Trans- actions on Graphics (TOG), 42(6):1–21, 2023

    Yitong Deng, Hong-Xing Yu, Diyang Zhang, Jiajun Wu, and Bo Zhu. Fluid simulation on neural flow maps.ACM Trans- actions on Graphics (TOG), 42(6):1–21, 2023

    2023

  8. [8]

    Scalarflow: a large-scale volumetric data set of real-world scalar transport flows for computer animation and machine learning.ACM Transactions on Graphics (TOG), 38(6):1–16, 2019

    Marie-Lena Eckert, Kiwon Um, and Nils Thuerey. Scalarflow: a large-scale volumetric data set of real-world scalar transport flows for computer animation and machine learning.ACM Transactions on Graphics (TOG), 38(6):1–16, 2019

    2019

  9. [9]

    Deblurgan: Blind motion deblurring using con- ditional adversarial networks

    Kupyn et al. Deblurgan: Blind motion deblurring using con- ditional adversarial networks. InCVPR, 2018

    2018

  10. [10]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Nah et al. Deep multi-scale convolutional neural network for dynamic scene deblurring. InCVPR, 2017

    2017

  11. [11]

    Fluidnexus: 3d fluid reconstruction and prediction from a single video

    Yue Gao, Hong-Xing Yu, Bo Zhu, and Jiajun Wu. Fluidnexus: 3d fluid reconstruction and prediction from a single video. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26091–26101, 2025

    2025

  12. [12]

    Neurofluid: Fluid dynamics grounding with particle- driven neural radiance fields

    Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle- driven neural radiance fields. InInternational Conference on Machine Learning, pages 7919–7929. PMLR, 2022

    2022

  13. [13]

    Single image haze removal using dark channel prior.IEEE transactions on pattern analysis and machine intelligence, 33(12):2341–2353, 2010

    Kaiming He, Jian Sun, and Xiaoou Tang. Single image haze removal using dark channel prior.IEEE transactions on pattern analysis and machine intelligence, 33(12):2341–2353, 2010

    2010

  14. [14]

    Φflow (PhiFlow): Differen- tiable simulations for pytorch, tensorflow and jax

    Philipp Holl and Nils Thuerey. Φflow (PhiFlow): Differen- tiable simulations for pytorch, tensorflow and jax. InInterna- tional Conference on Machine Learning. PMLR, 2024

    2024

  15. [15]

    Taichi: a language for high- performance computation on spatially sparse data structures

    Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan- Kelley, and Frédo Durand. Taichi: a language for high- performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6):201, 2019

    2019

  16. [16]

    Difftaichi: Differentiable programming for physical simulation.ICLR, 2020

    Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation.ICLR, 2020

    2020

  17. [17]

    Freeman, and Frédo Durand

    Yuanming Hu, Jiafeng Liu, Xuanda Yang, Mingkuan Xu, Ye Kuang, Weiwei Xu, Qiang Dai, William T. Freeman, and Frédo Durand. Quantaichi: A compiler for quantized simula- tions.ACM Transactions on Graphics (TOG), 40(4), 2021

    2021

  18. [18]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

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

    2023

  19. [19]

    Lagrangian neural style transfer for fluids.ACM Transactions on Graphics (TOG), 39(4):52–1, 2020

    Byungsoo Kim, Vinicius C Azevedo, Markus Gross, and Barbara Solenthaler. Lagrangian neural style transfer for fluids.ACM Transactions on Graphics (TOG), 39(4):52–1, 2020

    2020

  20. [20]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 4015–4026, 2023

    2023

  21. [21]

    Zero-1-to-3: Zero-shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. InProceedings of the IEEE/CVF international conference on computer vision, pages 9298– 9309, 2023

    2023

  22. [22]

    Position based fluids

    Miles Macklin and Matthias Müller. Position based fluids. ACM Transactions on Graphics (TOG), 32(4):1–12, 2013

    2013

  23. [23]

    Unified particle physics for real-time appli- cations.ACM Transactions on Graphics (TOG), 33(4):1–12, 2014

    Miles Macklin, Matthias Müller, Nuttapong Chentanez, and Tae-Yong Kim. Unified particle physics for real-time appli- cations.ACM Transactions on Graphics (TOG), 33(4):1–12, 2014

    2014

  24. [24]

    Particle- based fluid simulation for interactive applications

    Matthias Müller, David Charypar, and Markus Gross. Particle- based fluid simulation for interactive applications. InProceed- ings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation, pages 154–159, 2003

    2003

  25. [25]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

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

    Pith/arXiv arXiv 2022

  26. [26]

    Development and evaluation of gappy-pod as a data recon- struction technique for noisy piv measurements in gas turbine combustors.Experiments in Fluids, 57(7):122, 2016

    Pankaj Saini, Christoph M Arndt, and Adam M Steinberg. Development and evaluation of gappy-pod as a data recon- struction technique for noisy piv measurements in gas turbine combustors.Experiments in Fluids, 57(7):122, 2016

    2016

  27. [27]

    An unconditionally stable maccormack method.Journal of Scientific Computing, 35(2):350–371, 2008

    Andrew Selle, Ronald Fedkiw, Byungmoon Kim, Yingjie Liu, and Jarek Rossignac. An unconditionally stable maccormack method.Journal of Scientific Computing, 35(2):350–371, 2008

    2008

  28. [28]

    Aerial imagery pile burn detection using deep learning: The flame dataset

    Alireza Shamsoshoara, Fatemeh Afghah, Abolfazl Razi, Lim- ing Zheng, Peter Z Fulé, and Erik Blasch. Aerial imagery pile burn detection using deep learning: The flame dataset. Computer Networks, 193:108001, 2021

    2021

  29. [29]

    Vision transformers for single image dehazing.IEEE Transactions on Image Processing, 32:1927–1941, 2023

    Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing.IEEE Transactions on Image Processing, 32:1927–1941, 2023

    1927

  30. [30]

    John Wiley & Sons, 2013

    M’hamed Souli and David J Benson.Arbitrary Lagrangian Eulerian and fluid-structure interaction: numerical simula- tion. John Wiley & Sons, 2013. 9

    2013

  31. [31]

    Deep learning methods for reynolds-averaged navier–stokes simulations of airfoil flows.AIAA Journal, 58 (1):25–36, 2020

    Nils Thuerey, Konstantin Weißenow, Lukas Prantl, and Xi- angyu Hu. Deep learning methods for reynolds-averaged navier–stokes simulations of airfoil flows.AIAA Journal, 58 (1):25–36, 2020

    2020

  32. [32]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 20697–20709, 2024

    2024

  33. [33]

    Images speak in images: A generalist painter for in-context visual learning

    Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830–6839, 2023

    2023

  34. [34]

    Physics-based fluid simulation in computer graphics: Survey, research trends, and challenges

    Xiaokun Wang, Yanrui Xu, Sinuo Liu, Bo Ren, Jiri Kosinka, Alexandru C Telea, Jiamin Wang, Chongming Song, Jian Chang, Chenfeng Li, et al. Physics-based fluid simulation in computer graphics: Survey, research trends, and challenges. Computational Visual Media, pages 1–56, 2024

    2024

  35. [35]

    Structured 3d latents for scalable and versatile 3d generation.arXiv preprint arXiv:2412.01506, 2024

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation.arXiv preprint arXiv:2412.01506, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency.arXiv preprint arXiv:2407.17470, 2024

    Yiming Xie, Chun-Han Yao, Vikram V oleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency.arXiv preprint arXiv:2407.17470, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Sv4d 2.0: Enhancing spatio-temporal consistency in multi-view video diffusion for high-quality 4d generation.arXiv preprint arXiv:2503.16396, 2025

    Chun-Han Yao, Yiming Xie, Vikram V oleti, Huaizu Jiang, and Varun Jampani. Sv4d 2.0: Enhancing spatio-temporal consistency in multi-view video diffusion for high-quality 4d generation.arXiv preprint arXiv:2503.16396, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Inferring hybrid neural fluid fields from videos.Advances in Neural Information Processing Systems, 36, 2024

    Hong-Xing Yu, Yang Zheng, Yuan Gao, Yitong Deng, Bo Zhu, and Jiajun Wu. Inferring hybrid neural fluid fields from videos.Advances in Neural Information Processing Systems, 36, 2024

    2024

  39. [39]

    Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048, 2024

    Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Tomofluid: Reconstructing dynamic fluid from sparse view videos

    Guangming Zang, Ramzi Idoughi, Congli Wang, Anthony Bennett, Jianguo Du, Scott Skeen, William L Roberts, Peter Wonka, and Wolfgang Heidrich. Tomofluid: Reconstructing dynamic fluid from sparse view videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1870–1879, 2020

    2020

  41. [41]

    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. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  42. [42]

    Hunyuan3d 2.0: Scaling diffusion mod- els for high resolution textured 3d assets generation.arXiv preprint arXiv:2501.12202, 2025

    Zibo Zhao, Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang, Xianghui Yang, et al. Hunyuan3d 2.0: Scaling diffusion mod- els for high resolution textured 3d assets generation.arXiv preprint arXiv:2501.12202, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Vid2fluid: 3d dynamic fluid assets from single-view videos with generative gaussian splatting.arXiv preprint arXiv:2503.00868, 2025

    Zhiwei Zhao, Alan Zhao, Minchen Li, and Yixin Hu. Vid2fluid: 3d dynamic fluid assets from single-view videos with generative gaussian splatting.arXiv preprint arXiv:2503.00868, 2025

    arXiv 2025

  44. [44]

    Stable virtual camera: Gen- erative view synthesis with diffusion models.arXiv preprint arXiv:2503.14489, 2025

    Jensen Zhou, Hang Gao, Vikram V oleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rup- precht, and Varun Jampani. Stable virtual camera: Gen- erative view synthesis with diffusion models.arXiv preprint arXiv:2503.14489, 2025. 10 WildSmoke: Ready-to-Use Dynamic 3D Smoke Assets from a Single Video in the Wild Supplementary Material A...

    Pith/arXiv arXiv 2025