acceptodds
Withdrawn from ICLR 2027

VortMixture: Vortex-Driven Dynamic Smoke Reconstruction with Local Gaussian Mixtures

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reconstructing dynamic smoke from sparse multi-view observations is challenging due to its semi-transparent appearance, complex vortical motion, and severe ambiguity in recovering time-varying volumetric structure. Existing approaches commonly incorporate fluid priors as additional objectives alongside visual reconstruction losses, requiring heterogeneous physical and visual constraints to be jointly balanced. We propose VortMixture, a dynamic smoke reconstruction framework that embeds fluid structure directly into the motion representation while retaining flexible visual modeling. Specifically, we parameterize the Eulerian velocity field with vortex kernels, yielding divergence-free motion by construction and spatially coupling localized vorticity with the surrounding flow. The resulting field advects Lagrangian Gaussian primitives, while propagated vortex strengths model vortical evolution over time. To capture local variations beyond advection, we further associate each transported primitive with a local Gaussian mixture, increasing the capacity for density and appearance modeling without altering its underlying carrier trajectory. Experiments across three real-world smoke datasets demonstrate consistent improvements in sparse-view smoke reconstruction quality over prior methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dynamic Smoke Reconstruction, Physical simulation

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_McgzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on smoke reconstruction from sparse views via vortex-based motion representation and Gaussian primitives. Instead of an Eulerian velocity field on fixed grid, this paper advocates a vortex-velocity joint representation to better track vortex motion and guarantee the velocity divergence-free condition. Thereafter, the LGM module is presented to locally reallocate and refine the visual details. The reconstruction results are evaluated on FluidNexus and ScalarFlow, as well as physical validation on synthetic data.

Critical strengths

The introduction of vortex representation is interesting and practical in fluid dynamics, which especially makes sense for smoke phenomena.

The paper reports competitive performance compared to several SOTA methods, e.g., PICT, FluidNexus.

Critical weaknesses

The tangible benefit of the claimed contribution is still not clear to me. Please refer to Questions For Authors below.

Questions for authors
  1. In the abstract, the paper claims “existing approaches require physical and visual constraints to be balanced”, and the paper “embeds fluid structure into…”. However, after reading the entire text, the physics and vision balance does not differ fundamentally from mainstream methodology. Please clarify what is novel in this regard.
  2. Please explain why ”vortex kernels constrain a broader spatial domain” compared with Eulerian velocity on grid under sparse views.
  3. The paper did not discuss or compare with recent works, such as NeuSmoke, GauSmoke, SmokeSVD. A direct comparison with GauSmoke seems quite essential.
  4. Biot-Savart law is traditionally used to describe the magnetic field generated by a constant electric current. Please justify its applicability and numerical treatment in your smoke reconstruction setting.
  5. The vortex representation is employed in a quite direct manner, via transport with velocity field and advection. What is the main gain and advantage of the vortex representation over grid-based velocity fields?
  6. The initialization of the vortex field is insufficiently described for reproduction. Please provide details.
  7. As for LGM module, what is its difference from an adaptive or two-level 3DGS? Table 6 shows marginal advantages over a direct application.
  8. The design of L_reg appears somewhat ad hoc.
  9. Table 3 reports PSNR of 31.72, while Figure 6 shows 26.60?
  10. Are any constraints imposed to improve long-term temporal consistency?
Reviewer_kdwBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents VortMixture for dynamic smoke reconstruction from sparse multi-view videos. It parameterizes the velocity field using fixed-position, regularized Biot–Savart vortex kernels whose strengths are optimized per frame and regularized by a vortex-stretching-inspired update. The resulting divergence-free field transports Gaussian carriers. In a subsequent appearance-refinement stage, each carrier is represented by a local Gaussian mixture whose weighted center is constrained to follow the carrier trajectory. The paper reports improvements in novel-view rendering on three real-world datasets and evaluates recovered velocity on a synthetic sequence.

Critical strengths

The method achieves some improvements in novel-view rendering, and Table 6 includes a useful Gaussian-count-matched comparison of the local mixture against direct advection.

Critical weaknesses
  1. Limited novelty and insufficient motivation. The main components closely resemble existing methods, including HyFluid, FluidNexus, and LagrangianSplats. The paper does not clearly motivate the proposed modifications or establish their advantages over these approaches. See Question 1.

  2. Questionable physical consistency of vortex evolution. The vortex kernels remain fixed in space (an Eulerian representation), while their strengths are propagated using only vortex stretching, resembling a Lagrangian evolution without vorticity advection. This appears physically inconsistent and lacks adequate justification. See Question 2.

  3. Insufficient justification for local Gaussian mixtures. The additional appearance flexibility may compensate for inaccurate motion rather than improve physical reconstruction. The representation's necessity and benefits beyond increased fitting capacity are not convincingly demonstrated. See Question 3.

  4. Unclear and potentially unfair re-simulation evaluation. Re-simulation appears to allow additional appearance optimization, making it unclear whether the reported results are directly comparable to baselines or genuinely measure recovered motion accuracy. See Question 4.

  5. Insufficient experimental evaluation. LgSplats exhibits substantial apparent overfitting on the FluidNexus datasets but performs considerably better relative to competing methods on ScalarReal, raising concerns about dataset-dependent behavior and whether the comparisons fairly reflect methodological differences. Evaluation on additional datasets, particularly those originally used by LgSplats, is needed. Moreover, the absence of comparative videos makes it difficult to assess dynamic reconstruction quality during review, even if source code is available. See Question 5.

  6. Inconsistent velocity evaluation metrics. Several reported AAE and Cos-v values appear mathematically inconsistent with their definitions in Appendix D.1, raising concerns about the reliability of the velocity comparisons. See Question 6.

Questions for authors
  1. What is the precise relationship between the proposed vortex kernel and classical vortex-particle/regularized Biot–Savart representations? Why is this kernel preferable to existing divergence-free bases, including those used in Hyfluid and LagrangianSplats?

  2. How is Eq. (3) derived for fixed kernel centers? Where is vorticity advection represented, and how are the optimized coefficients related to the physical curl of the velocity? What happens when the propagation regularizer is removed?

  3. What additional physical or statistical benefit is gained from the local mixtures beyond greater fitting capacity?

  4. For the proposed method and baselines, what is optimized during re-simulation? Are target-frame observed images used at any stage? It would be nice to have a matched, pure-advection evaluation without target-frame appearance refinement.

  5. Dataset coverage, baseline generalization, and temporal evaluation. LgSplats achieves relatively high input-view PSNR but substantially lower novel-view PSNR on FluidNexus-Smoke (35.69 vs. 20.35 dB) and FluidNexus-Ball (34.14 vs. 19.08 dB), suggesting severe generalization difficulties. Yet its relative performance against competing methods is considerably better on ScalarReal. What explains this difference? Could it reflect dataset-specific behavior, implementation choices, or evaluation settings rather than an inherent advantage of the proposed method? I suggest including benchmarks originally evaluated by LgSplats to provide a more balanced comparison.

  6. The AAE and Cos-v values in Table 5 appear inconsistent with Eqs. (19)–(20). For the same evaluation samples, Cos-v ≥ 1 − AAE·π/180, which is violated by HyFluid and PICT. Could the authors clarify the discrepancy and verify the reported results?

  7. Could the authors provide synchronized video comparisons of complete reconstructed sequences? Source code alone is not a practical substitute for directly inspecting and comparing dynamic results during the review process.

Reviewer_sDqxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VortMixture proposes to separate motion and visual reconstruction of dynamic smoke. It first parameterizes a velocity field with vortex kernels and uses it to propagate gaussian primitives, capturing the underlying motion dynamics. To achieve higher-fidelity visual reconstruction, the gaussian primitives are associated with mixtures of gaussians for appearance optimization under the fixed dynamics. Experiments validate the accuracy of the recovered velocity field on synthetic data and improvements in dynamics and appearance reconstruction over baselines on real-world data.

Critical strengths
  • Using local Gaussian mixtures to better capture visual details and local deformations is straightforward and well motivated.
  • The analytical velocity-field representation appears effective, with improved velocity accuracy over baselines.
Critical weaknesses
  • The local gaussian mixtures do not seem to explicitly constrain individual component offsets to local neighborhoods. Could mixture components extend beyond their local carrier regions? This may weaken the connection between the rendered smoke and the underlying physical motion, as the mixtures could have enough flexibility to visually represent motion that differs from the recovered dynamics.
  • The re-simulation evaluation allows appearance refinement, which makes it harder to assess how well the reconstructed physics and appearance reproduce the observed dynamics without further optimization.
Questions for authors
  • Could the authors provide an analysis of component displacements within the local gaussian mixtures?
  • For re-simulation, could the authors provide results without appearance refinement? Are the baselines evaluated with or without appearance refinement allowed?
  • Would it be possible to extend the ablations to report results alongside the baselines in the main tables? This would provide a more explicit comparison of the improvements brought by the motion representation versus the additional degrees of freedom from appearance refinement.

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

Related papers

Discussion (0)

Sign in to comment.