acceptodds
Under review as a conference paper at ICLR 2027

VolSmoke: Reconstructing Smoke from a Single View with Density-Space Priors

Abstract

Recovering time-varying 3D smoke density and velocity from monocular video would enable accessible volumetric capture and physics-based re-simulation, but the problem is severely underconstrained: each pixel constrains only a ray integral, leaving the density distribution in space ambiguous. Recent methods address this by generating novel views, yet such image-space priors need not correspond to a consistent 3D volume, which can hinder subsequent physical reconstruction. We instead resolve the missing information directly in 3D density space, where a single volume provides a shared explanation across views and a physical state for transport. We introduce VolSmoke, a monocular reconstruction framework built around a view-conditioned latent prior over 3D smoke density. At inference, we sample plausible densities from this prior and use render-guided reverse diffusion to align each sample with the observed view. Temporal coordination further lets neighboring frames jointly constrain the otherwise ambiguous density. The resulting density sequence is then jointly refined with inflow source and a continuously divergence-free velocity field under rendering and transport constraints. On the real-world ScalarFlow dataset and a challenging synthetic Plume benchmark, VolSmoke substantially improves held-out-view reconstruction and velocity recovery over state-of-the-art novel-view baselines, leading to more accurate re-simulation. Our density prior can be trained from either synthetic simulations or simple tomographic reconstructions from sparse-view real-world captures, without greater data requirements than prior image-space approaches. These results show that density-space completion improves not only appearance reconstruction, but also the physical state required for velocity estimation and re-simulation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Fluid Capture, Single-View Reconstruction, Physical Priors

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1fALOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses 3D smoke reconstruction from a video taken with one fixed, calibrated camera. Instead of synthesizing missing camera views, it learns a prior over 3D density. An autoencoder represents the density in a monoplanar latent space, and an image-conditioned diffusion model learns to generate plausible latents. At test time, the frozen model is sampled with rendering feedback from the input view and constraints between nearby frames. A final, sequence-specific optimization fits density corrections, smoke inflow, and a divergence-free DFK velocity field using image and transport losses. The experiments examine novel-view rendering, velocity estimation, and re-simulation on Plume and ScalarFlow.

Critical strengths
  1. The Plume reconstruction is convincing in the shown sequence. The supplementary Plume15 clips show frames near 55, 74, and 93. The predicted plume keeps its overall head and stem in both the input and held-out views, including during cumulative replay. Several baselines shown in the same clips become much blurrier or lose the plume shape. This agrees with the reported Plume held-out PSNR: 34.10 dB for reconstruction and 34.92 dB for re-simulation, versus 30.04 and 28.73 dB for the strongest listed alternatives in Table 2. These videos are selected examples, and I have not independently reproduced the numbers.

  2. The paper makes a useful attempt to separate density quality from the physical refiner. The shared Stage-3 comparison reports a Plume velocity cosine of 0.496 for this method versus 0.188 for the best listed shared-refiner alternative. The ablations are informative: removing diffusion lowers held-out reconstruction PSNR from 34.10 to 25.00 dB, while removing temporal-window synchronization lowers it to 32.49 dB. These results support the role of the learned density prior and temporal coordination.

Critical weaknesses
  1. I cannot reconcile the ScalarFlow evaluation numbers. Tables 2 and 4 say they evaluate the recorded held-out cameras and report 32.99 PSNR / 0.956 SSIM / 0.081 LPIPS for Ours. Table 6 assigns exactly those values to 360-degree views rendered from reference volumes, but reports 33.80 / 0.9724 / 0.0789 for Ours in its recorded-camera row. The text in Appendix B.4 again says Tables 2 and 4 use recordings. Perhaps one table is mislabeled or contains values from a different run. As written, I cannot tell which evidence supports the main ScalarFlow comparison.

  2. The supplementary videos do not settle the broader physical claims. The ScalarFlow velocity and cumulative replay examples are labeled sim094, within the paper's sim010–103 training split. In the replay video, Ours is labeled 'Post-processed preview' and has no displayed PSNR, unlike the baselines. The crosswind example applies the same spatial warp to the rolled-out densities and the reference; it does not test a rollout under a new wind force. These demonstrations are useful, but they provide narrower evidence than a held-out ScalarFlow replay or a genuine force-change experiment would.

Questions for authors
  1. Which ScalarFlow results in Tables 2, 4, and 6 are measured against recorded camera images, and which are measured against rendered reference volumes? Why do 32.99 / 0.956 / 0.081 appear under both descriptions, while Table 6 gives another recorded-camera result of 33.80 / 0.9724 / 0.0789? Please provide the corrected mapping of numbers, scenes, cameras, frames, and reference images.

  2. Could the authors show an unprocessed reconstruction and cumulative replay from a held-out ScalarFlow scene (sim005–009) in the camera setting used for the main table? Please also explain the choice of sim094 for the current clips and what 'Post-processed preview' means for Ours.

  3. In the crosswind experiment, is wind introduced into the velocity/transport calculation, or is the stated warp applied after rollout? If it is the latter, the result would be better described as a warped replay than as evidence of response to an unseen force.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1fAL

We thank the reviewer for the careful and constructive evaluation. We sincerely apologize for the inconsistencies in our original presentation. We have corrected the tables, rerun the relevant evaluations, and updated the supplementary material.

  1. ScalarFlow evaluation and Tables 2, 4, and 6.

    We acknowledge errors in Table 6 and the FluidNexus entries in Tables 2 and 4. Our earlier FluidNexus evaluation used images rendered from multiview-reconstructed ScalarFlow volumes, which were mistakenly included in the recorded-camera comparison.

    We have rerun FluidNexus using its official ScalarFlow checkpoint, original recorded videos, and matched input/test cameras. The evaluation uses held-out scenes sim005--009, frames 50--149, with calib3 as input and calib1/2/4/5 as test views. FluidNexus achieves 31.36 dB PSNR / 0.9559 SSIM / 0.1402 LPIPS, compared with 32.99 / 0.956 / 0.081 for VolSmoke. These FluidNexus scores are our reproduced results, not values quoted from the original paper.

    We also corrected Table 6, where our recorded-camera and full-orbit reference-render results were inadvertently placed in the wrong rows. The corrected recorded-camera result (32.99 / 0.956 / 0.081) now agrees with Tables 2 and 4. The full-orbit reference-render setting is reported separately (32.87 / 0.950 / 0.088), with both protocols clarified in Appendix B.4. They remain substantially better than FluidNexus's 29.63 / .882 / .151, including a 3.24 dB PSNR advantage. We have corrected the reference-render entries and clarified both protocols in the manuscript.

    We sincerely apologize for the earlier evaluation inconsistencies and appreciate the reviewer for identifying them. After correcting the protocols and rerunning FluidNexus, the revised results continue to support the effectiveness of VolSmoke on held-out real-world ScalarFlow data.

  2. ScalarFlow supplementary videos and "Post-processed preview."

    We inadvertently included the training scene sim094 in our earlier supplementary examples. Both the cumulative re-simulation video and velocity-field visualization have now been replaced with results from the held-out test scene sim006. Our quantitative tables use the official ScalarFlow test-scene split.

    The label "Post-processed preview" referred to the final output after Stage 3 physical refinement, which is an integral part of our standard reconstruction pipeline, rather than an additional post-processing procedure. We have removed this ambiguous label and updated the PSNR annotations in the re-simulation video.

    These revised videos provide direct visual evidence of velocity recovery and cumulative transport on a genuinely held-out ScalarFlow scene.

  3. Crosswind re-simulation.
    The crosswind perturbation is implemented as a lateral density displacement applied after every advection step. Each sequence is rolled out from its initial density, and the perturbed density becomes the input to the next advection step. The GT density is independently rolled out using its reference transport under the same perturbation schedule.
    Thus, each perturbation influences subsequent transport, rather than being applied independently to frames of an already completed rollout. The low error relative to the correspondingly perturbed GT rollout demonstrates that our reconstructed fields can sustain accurate re-simulation under additional perturbations. We have clarified this procedure in the manuscript and supplementary material.

We sincerely appreciate the reviewer's detailed examination, which helped us identify and correct these issues. The corrected evaluations and new held-out supplementary results strengthen the experimental evidence while preserving our main conclusions regarding density reconstruction, velocity recovery, and transport consistency.

We hope these revisions address the reviewer's concerns about the reliability of our evaluation and would welcome any further questions.

Thank you for the detailed response and the substantial revisions. The authors have corrected the original ScalarFlow protocol description, rerun the FluidNexus baseline with the official ScalarFlow checkpoint and recorded input videos, replaced the training-scene supplementary example with a held-out scene, and clarified the stepwise crosswind re-simulation procedure.

These changes substantially improve the submission. The Plume results, ground-truth velocity evaluation, shared Stage-3 comparison, and ablations continue to support the value of density-space completion for monocular smoke reconstruction. The new ScalarFlow loss ablations also provide useful evidence that transport, mass, and vorticity terms contribute on real captured smoke. The crosswind experiment is now clearer: the density perturbation is applied after each advection step and is propagated into the next step, rather than being applied only to already completed frames.

However, I still have a concern about the final consistency of the central ScalarFlow results. Tables 2, 4, and 6 report the Ours test-view reconstruction metrics as 32.99 dB / 0.956 SSIM / 0.081 LPIPS, whereas the corrected ScalarFlow full-pipeline row in Table 7 reports 32.99 dB / 0.948 SSIM / 0.087 LPIPS. If the Table 7 values are the corrected final evaluation, the method does not outperform FluidNexus on SSIM (0.948 versus 0.9559), despite the broader claim of leading on all metrics. The PSNR and LPIPS results remain strong, but the manuscript should not leave these two final metric sets inconsistent.

The revised crosswind experiment strengthens the evidence for transporting a prescribed density edit through the recovered field, but it is not a test of a newly inferred velocity response to an unseen external wind force. The velocity field is still fitted within the observed interval, and the perturbation is a prescribed density displacement applied during rollout. This result should therefore be interpreted as evidence of robustness under a controlled downstream density edit, not as validation of unknown-force dynamics.

Overall, the authors have made meaningful progress and the method has clear technical value. Nevertheless, the unresolved inconsistency in the central ScalarFlow evaluation affects the interpretation of the main real-data comparison. I therefore maintain my recommendation of weak rejection (2), with confidence 3. This is a borderline recommendation rather than a clear rejection: the Plume evidence and the corrected parts of the ScalarFlow evaluation are promising, but the final quantitative claims should be internally consistent before acceptance.

We sincerely thank the reviewer for their exceptionally careful and patient assessment throughout the rebuttal process, and we apologize for this inconsistency and the oversight in our revisions.

During the rebuttal period, we reran the full set of ablation experiments on a different machine with the same hardware configuration, with the results given in Table 7. To ensure consistency, we also reran our complete pipeline, which produced slightly different numerical results from those in our original submission. Unfortunately, we acknowledge that we were not sufficiently careful in preparing our rebuttal revisions and failed to synchronize these updates across all tables. We apologize for this oversight and sincerely appreciate the reviewer's careful attention in identifying it.

We have chosen to retain the latest results, despite their slightly lower scores, and have updated Tables 2, 4, and 6 to be consistent with Table 7. The corrected ScalarFlow results are:

Method / versionPSNR ↑SSIM ↑LPIPS ↓
Ours — updated32.99.948.087
FluidNexus — reproduced31.36.956.140
Ours — originally reported32.99.956.081

We fully acknowledge that our updated SSIM is lower than FluidNexus, while our PSNR and LPIPS remain better. We have made an initial round of revisions throughout the manuscript to reflect these corrected results and clarify our claims. However, we recognize that this revision during the rebuttal may not have been sufficiently thorough yet. In the next revision, we will carefully review the entire manuscript word by word, verify the consistency of all reported results, and further narrow our claims to ensure they are fully supported by the evidence.

We also appreciate the reviewer's clarification regarding the crosswind experiment. We agree and apologize for the imprecise wording. The test scene is unseen during training, whereas the external crosswind force is known and prescribed, with its effect represented by a density-displacement correction during rollout. The recovered velocity remains fitted to the observed interval; we do not infer an unknown force or a new velocity response. The results support transporting this prescribed perturbation through the recovered field on the held-out test scene, without establishing unknown-force dynamics. We have made an initial clarification and will further refine the description to avoid overstating the implications of this experiment.

We are genuinely grateful for the reviewer's patience, thoroughness, and constructive criticism. Their careful feedback has helped us identify important shortcomings in our revisions and improve the accuracy and clarity of our submission.

Reviewer_MMrxOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VolSmoke, a three-stage framework combining conditional density-space diffusion, temporal coordination, and physics-based refinement for monocular smoke reconstruction. Experiments on ScalarFlow and Plume show improved novel-view rendering and velocity recovery. However, the methodological novelty and physical validity of the recovered fields remain insufficiently established.

Critical strengths

The paper addresses an important ill-posed reconstruction problem and provides a coherent pipeline connecting volumetric appearance with motion estimation.

The extensive ablations provide useful evidence for the contributions of individual pipeline components.

Critical weaknesses

Limited methodological novelty. The approach combines existing latent diffusion, rendering-based posterior guidance, and physics-based refinement. The paper does not sufficiently establish a fundamentally new inference principle beyond their integration.

Unreliable physical identifiability. Divergence-free parameterization and approximate transport losses do not uniquely recover velocity from monocular density observations. Moreover, the Plume prior uses different conditioning semantics during training and inference, raising concerns about distribution shift and the validity of the learned prior.

Questions for authors
  1. How does the method resolve the fundamental ambiguity between different velocity fields that explain similar density evolution?

  2. Why does the Plume diffusion prior remain reliable when its training and inference conditions differ in semantics?

1 comment
AuthorsOpenReview ↗

Response to Reviewer MMrx

We thank the reviewer for the feedback. We respectfully disagree with the characterization of our methodological contributions and clarify the novelty, physical accuracy, and conditioning protocol below.

  1. Methodological Novelty

    To our knowledge, VolSmoke is the first framework to recover dynamic smoke density and velocity from monocular video using a learned 3D density diffusion prior, addressing monocular ambiguity directly in density space rather than synthesizing missing views. Unlike FluidNexus and SmokeSVD, which rely on potentially inconsistent generated views, our prior produces a shared 3D volume consistent across viewpoints.

    More importantly, our key contribution is the mutual reinforcement of density inference and velocity recovery: improved density estimates better constrain velocity, while physically constrained velocity feeds back to refine density. Unlike Cloudy's static volumetric reconstruction, VolSmoke establishes this coupling for dynamic smoke.

    Experiments validate both directions: under an identical Stage 3 refiner, our density initialization achieves 0.496 velocity cosine versus 0.188 for SmokeSVD (Table 1); conversely, joint refinement reduces density MSE from 0.000621 to 0.000512 (Table 9). Removing temporal synchronization further lowers velocity cosine to 0.255 (Table 3).

    These results demonstrate a distinct density--velocity reconstruction framework, rather than a mere combination of existing modules.

  2. Velocity Ambiguity and Physical Validity

    Non-uniqueness is an inherent challenge of monocular reconstruction, not a limitation specific to our method. Our goal is not to guarantee a mathematically unique solution, but to recover a physically plausible and accurate 4D state from severely underconstrained observations.

    Beyond the learned density prior, our recovered velocities are jointly constrained by incompressibility, forward/backward density transport, mass consistency, recurrent rollout, and approximate vorticity dynamics. These constraints couple velocity fields across time and enforce both observed transport and physical consistency.

    Crucially, our physical accuracy is directly validated against ground-truth velocity, rather than only inferred from low divergence. On Plume, VolSmoke achieves a velocity cosine similarity of 0.496, substantially outperforming existing baselines, including SmokeSVD (0.061) and PINF (−0.004). Even under the same physical refiner, the strongest compared baseline reaches only 0.188 (Table 1). Our method also achieves the lowest velocity MSE among the compared methods.

    The physical constraints contribute directly to this accuracy: removing recurrent rollout reduces velocity cosine from 0.496 to 0.201, while removing vorticity regularization reduces it to 0.450 (Table 7). Cumulative replay further validates the recovered flow's ability to sustain accurate density transport over the observed interval.

    These substantial improvements in ground-truth velocity accuracy and transport consistency are central experimental contributions of our work, despite the fundamental non-identifiability of monocular reconstruction.

  3. Conditioning Semantics and Robustness

    We acknowledge the conditioning distribution shift between training and inference.

    For the matched image-conditioning configuration, training conditions are extracted from rendered images of training densities, while inference conditions are extracted from the observed images:

    \qquad C_{\mathrm{test}}=F(I_{\mathrm{test}}).$$ Both represent the same image-space quantities under the same conditioning operator. The difference is the source of the images, rather than their conditioning semantics; test-time density is not required. **Importantly, the strong results despite this shift demonstrate the empirical robustness of our framework.** On Plume, VolSmoke achieves **34.10 dB held-out-view PSNR versus 30.04 dB for the strongest baseline**, alongside substantially improved velocity recovery. Render-guided sampling, temporal synchronization, and physical refinement further align the reconstructed state with observations. These results demonstrate that the conditioning difference does not prevent accurate monocular reconstruction. We will clarify this distinction and the preprocessing protocol in Appendix A.3.
Reviewer_HgdYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper reconstructs 3D smoke density and velocity from a single video. It learns a diffusion prior over 3D density and samples a density for each frame, guided by rendering and temporal consistency. It then refines density, inflow, and a divergence-free velocity with transport losses. It reports better performance over current methods on ScalarFlow and a synthetic Plume dataset.

Critical strengths
  1. The problem is meaningful. Putting the prior on 3D density makes the result consistent across views by construction.
  2. The physics backbone is reasonable: divergence-free velocity and semi-Lagrangian transport with inflow. The physics-informed losses are sound.
  3. Extensive experiments and reasonable ablation for the losses used.
  4. The visual results are good.
Critical weaknesses
  1. L_vor (Eq. 14) assumes vorticity is only advected and stretched, never created. In a rising plume, vorticity mainly comes from buoyancy, so this term penalizes the main driver of the flow. The paper notes that buoyancy is omitted (App. A.5) but does not discuss the effect.
  2. Velocity has no real dynamics. Each transition has its own DFK field, and consecutive fields are linked only by the soft L_vor term. Divergence is also not a useful metric here, since DFK is divergence-free by construction. PINF, for example, has the same divergence (.00001) with a cosine of −.004. Vorticity error or prediction beyond the fitted interval may tell more about physical accuracy.
  3. The reproduced FluidNexus results are far below the published ones, both for visual and numerical. Especially for FluidNexus results on ScalarFlow, if pretrained ckpts from FluidNexus are used, novel view results should not be that bad.
  4. The loss ablations are on Plume only, and Table 7 reports reconstruction only on the input view. ScalarFlow has no loss ablation, and the margin there is only about 1.35 dB. It is therefore unclear whether the weaker terms, such as L_vor and L_mass, still help on real data.
Questions for authors
  1. App. B.4 indicates that the ScalarFlow results in Tables 2 and 4 use the recorded camera images but Table 6 labels the same numbers as “reference renders / 360°” and gives different numbers for the recorded setting (33.80/.9724/.0789). Please clarify. What settings do Tables 2 and 4 use, and how do the frames and scenes differ from the FluidNexus paper? For Plume, which FluidNexus modules were fine-tuned, on what data, and with what settings?
  2. Did you run the full FluidNexus pipeline, including novel-view video generation? If the official checkpoints are run end to end, do the ScalarFlow and Plume results change?
  3. Any other metrics that help show good physical properties?
4 comments
AuthorsOpenReview ↗

Response to Reviewer HgdY

We noticed that Reviewer HgdY's review appears to have been submitted to our paper by mistake. We have located the review intended for our submission under Submission #11408 (Reviewer puYo, https://openreview.net/forum?id=BogMmyGKy9), with a rating of 3 (Weak Accept). We therefore provide our point-by-point rebuttal to this review below.

  1. Buoyancy-induced vorticity in Eq. (14).
    We agree that L_vor omits buoyancy-induced vorticity generation and may penalize it. However, density transport and rendering provide the primary constraints on recovered motion. With inflow accounted for, the velocity must explain the observed density evolution, thereby implicitly capturing the effects of buoyancy on smoke motion. The vorticity loss serves as a soft regularizer under the inviscid, unforced approximation, favoring transport-consistent velocities with smaller unmodeled vorticity-source residuals. This observation-driven formulation is consistent with prior work on multi-view video reconstructionpin such as PINF (Chu et al., 2022), which uses simplified physics to regularize underdetermined reconstruction and seeks "a possible solution with minimal influence from extra forces, pressure difference, and viscosity."(PINF, Sec. 3.2).
    Explicit buoyancy modeling would introduce additional unknowns, such as temperature or buoyancy fields, that are difficult to infer from monocular video. Removing L_vor reduces held-out reconstruction PSNR by 0.87 dB on ScalarFlow, supporting its practical benefit despite the approximation. Appendix A.5 now clarifies its role and limitations.

  2. Coupling between DFK fields and evaluation of velocity accuracy.
    Although each transition has separate DFK parameters, the velocities are jointly constrained by forward density transport, backward cycle consistency, and recurrent rollout, coupling successive fields through short- and long-term advection. The vorticity loss further constrains their temporal evolution through vorticity advection and stretching. Together with analytical incompressibility, these objectives encourage a physically consistent 4D flow without an explicit differentiable Navier--Stokes solver.
    Our method achieves the best velocity accuracy against Plume ground truth in both cosine similarity and MSE (Table 1), while cumulative density replay on held-out views further validates transport consistency. We additionally report vorticity-evolution residuals in Appendix B.9 (Table 10), evaluated on 46 adjacent frame pairs under an inviscid, unforced approximation. Ours achieves 1.91 × 10⁻³, outperforming FluidNexus, HyFluid, and PINF, with GT at 8.49 × 10⁻⁴. SmokeSVD obtains a lower residual (7.44 × 10⁻⁴), but reconstructs a much simpler velocity field that deviates substantially from GT. In contrast, our method achieves both the highest fidelity to GT velocity and strong consistency with the underlying vorticity dynamics.

  3. FluidNexus reproduction and the ScalarFlow evaluation protocol. We apologize for the inconsistencies in our original evaluation. The recorded-camera comparison mistakenly included a FluidNexus result obtained using input video rendered from ScalarFlow reference volumes.
    We reran FluidNexus using its officially released ScalarFlow checkpoint, the original recordings, and the same calibrated input and test cameras. The recorded-camera comparison uses sim005--009, frames 50--149, with calib3 as input and calib1/2/4/5 held out. FluidNexus achieves 31.36 PSNR / 0.9559 SSIM / 0.1402 LPIPS, compared with 32.99 / 0.956 / 0.081 for VolSmoke.
    We also corrected Table 6, where recorded-camera and full-orbit reference-render results were inadvertently swapped. The revised recorded-camera results now match Tables 2 and 4, and Appendix B.4 clearly distinguishes the evaluation protocols.

  4. ScalarFlow loss ablations and held-out-view reconstruction. We extended Table 7 with all eight loss-group ablations on ScalarFlow, including held-out-view reconstruction and replay metrics. Removing the mass and vorticity regularizers individually lowers held-out reconstruction PSNR by 0.68 and 0.87 dB, respectively, and held-out replay PSNR by 0.72 and 0.45 dB. Both removals also worsen held-out SSIM and LPIPS and reduce velocity cosine similarity against the five-camera reconstruction reference. These results confirm that both regularizers contribute to reconstruction quality and transport consistency on real captured smoke.

We sincerely thank the reviewer for the positive assessment and constructive feedback. We hope that the clarified physical formulation, corrected baseline evaluation, and new ScalarFlow ablations address the main concerns and would welcome any further questions.

1 Fig. 4 vs. Table 4. Table 4 gives FluidNexus 31.36 dB but Fig. 4 still shows great noise. Was the figure regenerated from the corrected run?
2 Was FluidNexus run end to end, including novel-view video generation, with the official ScalarFlow checkpoint?
3 Which FluidNexus modules were fine-tuned, on what data, and with what settings? The re-simulations look nearly empty.
4 Table 7. ScalarFlow “Full pipeline” has identical Given and Test SSIM/LPIPS (.956/.081).

AuthorsOpenReview ↗

Response to Reviewer HgdY

Thank you for identifying these inconsistencies. We apologize for the oversights and address each point below.

1. Fig. 4 vs. Table 4.
You are correct: we updated the quantitative results but forgot to replace the earlier FluidNexus visualization in Fig. 4. We have now updated both the ScalarFlow reconstruction and re-simulation panels using the latest reproduction. The reconstruction panels show 34.51/32.08 dB for the given/test views, and the re-simulation panels show 33.231/31.255 dB. These are displayed-frame scores; 31.36 dB in Table 4 is the aggregate reconstruction score. The updated panels and labels are highlighted in yellow.

2. Complete FluidNexus pipeline and official checkpoints.
Yes. Our latest ScalarFlow reproduction includes Zero123 novel-view generation and CogVideoX video generation before reconstruction, using the official ScalarFlow-specific checkpoints from yuegao/FluidNexusModels:

  • Zero123: 2024-04-24T14-04-16_scalar_flow/checkpoints/step=000015499.ckpt. Strict loading succeeded, and all four target cameras (00/01/03/04) completed frames 20–160.
  • CogVideoX: lora_cogvidx5b_scalar_flow_i2v_fix3samples-09-09-05-39/10000/mp_rank_00_model_states.pt. All four cameras completed three 65-frame windows starting at frames 20, 76, and 131.

The generated sequences were then used for FluidNexus reconstruction.

3. Fine-tuning and nearly empty re-simulations.
We did not fine-tune any FluidNexus module on ScalarFlow. “ScalarFlow-fine-tuned” refers to the released weights that the FluidNexus authors fine-tuned on ScalarFlow. We used these weights directly, without additional ScalarFlow fine-tuning data or training settings. Our additional fine-tuning was confined to Plume and was not used for these ScalarFlow results.

The nearly empty supplementary clip came from an earlier reproduction that we had overlooked when updating the results. We apologize for this omission. The supplementary HTML now contains the latest FluidNexus re-simulation video, and the corresponding images in Fig. 4(b) have also been updated.

4. Identical Given/Test SSIM and LPIPS in Table 7.
This was an oversight: the Test entries incorrectly repeated the Given-view values instead of the latest experimental results. We have corrected the ScalarFlow “Full pipeline” row: Given SSIM/LPIPS = .956/.081; Test SSIM/LPIPS = .948/.087. The corrections are highlighted in yellow.

Thank you again for your careful review.

AuthorsOpenReview ↗

Response to Reviewer HgdY

Further Clarification and Appreciation

We would like to sincerely thank the reviewer again for their exceptionally careful and patient examination of our results, figures, and supplementary materials. Your detailed feedback has been extremely valuable in helping us identify inconsistencies that we had overlooked.

Following your response and an observation from Reviewer 1fAL, we realized that the corrected ScalarFlow results in Table 7 were not synchronized with Tables 2, 4, and 6. During the rebuttal, we reran our complete pipeline on a different machine with the same hardware configuration to maintain consistency with the updated ablations. The resulting scores differed slightly from those originally reported, but we were not sufficiently careful in propagating these updates throughout the manuscript.

We have now updated Tables 2, 4, and 6 to consistently report the latest test-view results (in Table 7). The corrected ScalarFlow results are:

Method / versionPSNR ↑SSIM ↑LPIPS ↓
Ours — updated32.99.948.087
FluidNexus — reproduced31.36.956.140
Ours — originally reported32.99.956.081

We have chosen to retain the latest results, despite their slightly lower scores, and acknowledge that our updated SSIM is slightly lower than FluidNexus (0.948 vs. 0.956), while PSNR and LPIPS remain better.

We have made an initial round of corrections to the manuscript and its claims. However, we recognize that our revisions during the rebuttal were not sufficiently thorough. In the next revision, we will carefully review the entire manuscript, verify the consistency of the reported results and visualizations, and further narrow our claims to accurately reflect the experimental evidence.

We sincerely apologize for these oversights and are especially grateful for the reviewer's patience, thoroughness, and constructive feedback throughout this process. Your careful examination has helped us substantially improve the reliability and clarity of our submission.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Cover Letter

We thank the reviewers for their constructive feedback and for recognizing the importance of monocular smoke reconstruction, the quality of the Plume results, and the informative ablations. We have submitted a revised manuscript (PDF) and an updated supplementary webpage (HTML), with changes highlighted in yellow. The main revisions are summarized below.

  1. Velocity recovery and physical validity (HgdY, MMrx, 1fAL).
    Monocular smoke reconstruction is inherently underdetermined: multiple density and velocity fields can explain the same observations. Rather than claiming unique recovery or constructing a predictive physics solver, our goal is to recover a physically plausible, observation-consistent 4D state.

    Stage 2 reduces this ambiguity by combining a learned 3D density prior with rendering guidance and temporal synchronization, producing coherent density sequences that better constrain velocity recovery. On Plume, removing temporal synchronization reduces velocity cosine similarity from 0.496 to 0.255 (Table 3).

    Stage 3 further improves physical consistency by jointly refining density and divergence-free velocity. Rendering maintains observation consistency, while forward transport, cycle consistency, and recurrent rollout constrain the recovered velocity. Through this coupling, the effects of buoyancy on the observed density evolution can inform the recovered motion. The vorticity loss complements transport by penalizing departures from advection and stretching under the inviscid, unforced incompressible vorticity equation. Together, these constraints favor flows that explain observed transport with smaller unmodeled vorticity sources. Appendix A.5 clarifies this interpretation and the limitation of omitting explicit buoyancy-driven vorticity generation.

    Joint refinement reduces Plume density MSE from 0.000621 to 0.000512 (Table 9), while removing the vorticity loss lowers velocity cosine similarity from 0.496 to 0.450 (Table 7). Cumulative replay on held-out views further evaluates transport consistency over the observed interval.

  2. Corrected ScalarFlow evaluation and expanded ablations (HgdY, 1fAL).
    The original recorded-camera comparison mistakenly included a FluidNexus result obtained using an input video rendered from ScalarFlow reference volumes. We reran FluidNexus using its official ScalarFlow checkpoint, the original recordings, and the same calibrated input and test cameras as VolSmoke. The rerun achieves 31.36 dB PSNR, 0.956 SSIM, and 0.140 LPIPS-Alex. The original Table 6 also placed our recorded-camera and reference-render results in the opposite rows. The revised Table 6 distinguishes the two settings, and its recorded-camera results now agree with Tables 2 and 4. Table 7 adds eight ScalarFlow loss-group ablations and held-out-view reconstruction metrics for both datasets. Removing the mass and vorticity regularizers individually lowers ScalarFlow held-out reconstruction PSNR by 0.68 and 0.87 dB, respectively, and replay PSNR by 0.72 and 0.45 dB, supporting their benefits on real data.

  3. Updated supplementary HTML and re-simulation descriptions (1fAL).
    The original supplement inadvertently included a clip from sim094, a training scene. The updated supplement includes results for the held-out ScalarFlow scene sim006 and labels our reconstructions as the final outputs of Stage 3. We also clarified the crosswind experiment in Section 4.3 and Figure 5. Starting from the first-frame density, we apply a prescribed lateral displacement to the density after each advection step. The recovered velocity transports the edited density in the next step, and the reference rollout follows the same perturbation schedule. This demonstrates how the recovered flow can carry prescribed density edits forward during re-simulation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.