acceptodds
Withdrawn from ICLR 2027

VHeat-UDF: Regularizing Neural Unsigned Distance Fields with Vector Heat Diffusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

We present VHeat-UDF, a method for learning unsigned distance fields regularized by vector heat diffusion. Given potentially inconsistent surface normals, vector heat diffusion propagates geometric orientation into the ambient space, providing effective regularization for UDF learning. To address singular behavior near field singularities, we introduce a rescaled vector Dirichlet energy that remains well-defined in such regions. VHeat-UDF provides accurate distance queries and robust surface reconstruction, even from incomplete or sparse inputs. Experiments comparing our approach with existing methods and scalar heat diffusion-based regularization demonstrate its effectiveness and robustness. Code will be released upon acceptance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Neural unsigned distance fields, vector heat diffusion, implicit geometric representation.

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_PbBxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper regularises neural UDF learning with a "vector heat diffusion" term applied only early in training. This term reduces to a Hessian energy normalised by the mean gradient norm. It helps the field converge at the start of training, but novelty over existing Hessian regularisers is limited and not discussed.

Critical strengths

The regularisation seems to stabilise training and improve convergence.

Critical weaknesses
  • The proposed regulariser reduces to Hessian regularisation.

    • Proposition 2 cancels the normalisation: v=C∇fv = C\nabla f (Eq. 19), so the smoothness term is the squared Frobenius norm of the Hessian, up to a global factor C2C^2 (Eq. 21). Since CC depends on ff, this factor also changes the effective weight of the term during training. Why choose v=a(x)∇f^(x)v = a(x)\hat{\nabla f}(x) rather than simply A∇f(x)A\nabla f(x) with a constant AA?
    • Hessian regularisation is well studied: ∣tr(H)∣|\text{tr}(H)| in DiGS [1], which is also annealed to zero during training; ∣det⁡(H)∣|\det(H)| in NSH [2]; and the Monge–Ampère term in S²DF [3]. These works are cited in the related work, but the proposed term is never related to them. The authors should acknowledge this equivalence explicitly and compare against these terms under the same schedule.
  • The motivation for vector heat diffusion is unclear.

    • In vector heat diffusion, the source u0=n\mathbf{u}_0 = n is imposed on SS and diffused into the volume. Here, the diffused field is u=a∇f^\mathbf{u} = a\hat{\nabla f}, which cannot match this source on SS: ∇f\nabla f is undefined there for a UDF, and for a smooth ff attaining its minimum on SS (Eq. 14), ∇f=0\nabla f = 0, so the alignment term ∫S(1−∣∇f^⋅n∣)2\int_S (1 - |\hat{\nabla f} \cdot n|)^2 is constant. The boundary condition of the diffusion is therefore ill-posed. Together with the previous point, it is unclear what the vector heat diffusion formulation adds beyond a Hessian penalty.
  • Proposition 1 is weak.

    • It assumes the symmetry of a UDF rather than proving it: that ff is double-sided around SS follows from the definition of a UDF (f(p±rn)=rf(p \pm r n) = r near a smooth point p∈Sp \in S), but not from the stated hypotheses f≥cf \ge c, f∣S=cf|_S = c. Why not state this directly in terms of the UDF definition, and why introduce a general lower bound cc instead of 00?
    • The invariance under n↦−nn \mapsto -n follows immediately from the absolute value in Eq. 3, so the random-flip experiment in C.3.1 verifies a property that holds by construction.
  • Missing ablations. The paper should ablate LvdsmL_{vd}^{sm}, LvdbnL_{vd}^{bn}, λ\lambda, ε\varepsilon, the schedule, and the chosen hyperparameters.

  • Errors in references.

    • For "Neural unsigned distance fields for implicit function learning." (l. 521), the following authors are given "Julian Chibane, Gerard Pons-Moll, et al." while the true authors are "Julian Chibane, Aymen Mir, Gerard Pons-Moll", missing Aymen Mir.
    • For HSDF (l.643), the following authors are given "Li Wang, Weikai Chen, Xiaoxu Meng, Bo Yang, Jintao Li, Lin Gao, et al.", while the true authors are "Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, Jintao Li, Lin Gao", missing Jie Yang.
    • For “Voronoi-assisted diffusion for computing unsigned distance fields from unoriented points” (l. 572), the arXiv entry (2510.12524) has since been retitled “Voronoi-Assisted Optimization for Diffusing Unsigned Distance Fields from Unoriented Points”, the reference should be updated.
Questions for authors
  • Could Table 1 report metrics on the whole dataset? Baselines fail on simple garments (e.g. S²DF in Fig. 8), whereas [3] reports near-perfect F-scores for S²DF, LevelSetUDF and CAP-UDF on 20 MGN shapes. What is the input density, and was Algorithm 1 applied to S²DF's squared output?
  • How is QQ chosen? Is CC computed per batch, and is it detached from the graph?
  • No runtime or memory usage is reported, although second-order derivatives are computed at every off-surface sample.

[1] DiGS
[2] NSH
[3] S²DF

Reviewer_kVbKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses spatial distance accuracy and surface reconstruction quality in neural unsigned distance fields (UDFs). It proposes a regularizer that encourages smoothness in a rescaled gradient field, enabling the UDF to maintain a more reasonable distance representation even in regions far from the surface. The paper focuses on rescaling the gradient field and using it for regularization, which is applied only during the early stages of training.

Critical strengths
  1. Distinguishing accurate spatial distance prediction from merely reconstructing a plausible zero level set is meaningful. The paper explicitly targets the former, which is potentially valuable for applications requiring reliable distance queries.
  2. Selected narrow-structure, incomplete-scan, and complex-interior examples show encouraging distance fields or reconstruction results.
Critical weaknesses
  1. Unclear incremental contribution over prior work. The advance over VAD and existing second-order regularization is insufficiently established. Under the normalization used in Appendix B.2, the proposed smoothness term reduces to a globally normalized squared-Hessian energy. Controlled comparisons with standard Hessian and related second-order penalties, together with ablations isolating normalization, normal supervision, and the training schedule, are needed to identify the source of improvement.

  2. Inconsistent normalization definitions. Eq. (3) normalizes gradients by ∥∇f∥+ε\|\nabla f\|+\varepsilon, whereas Appendix B.2 uses ∥∇f∥\|\nabla f\|. This discrepancy affects the derivative calculation and claimed Hessian equivalence, particularly in the low-gradient regions central to the method's motivation. The authors should reconcile these definitions and clarify which theoretical claims hold for the implemented regularizer.

  3. Limited systematic quantitative comparisons. Quantitative comparisons mainly cover a small number of objects. The largest experiment, on 50 ABC models, reports only the proposed method's results, without baselines on the same models. Larger baseline errors do not justify omitting their numerical results. Matched comparisons, including separate near-surface and off-surface distance errors, are needed to establish consistent improvements.

  4. Surface metrics do not directly evaluate the reconstructed surfaces. CD is computed from points obtained through gradient-based search (Algorithm 1), rather than samples from the final reconstructed surfaces. This measures point-localization accuracy but does not establish final reconstruction quality. The fairness of the comparison remains unclear without specifying equivalent evaluation procedures for the baselines. Metrics sampled from the actual output surfaces would provide more direct evidence.

  5. Unclear surface extraction and insufficient near-surface analysis. The paper does not clearly specify how the displayed surfaces are extracted or rendered from the UDF. Moreover, the existing illustrations do not sufficiently quantify near-surface minima, surface-location bias, or gradient reversal. Given the smooth network and second-order regularization, a local analysis across the surface is needed to establish whether off-surface improvements also preserve the UDF's near-surface behavior.

Questions for authors

Which normalization is implemented: Eq. (3) or Appendix B.2? Do the derivative simplifications and the stability and scale-invariance claims hold for that implementation with positive ε?

How does the implemented regularizer relate to globally normalized Hessian energy? Could you compare it with standard Hessian penalties under matched settings and isolate the effects of normalization, normal supervision, and training schedule?

How are the displayed surfaces extracted or rendered from the UDF? If they are meshes, could you report metrics sampled from those meshes and clarify the extraction and evaluation procedures used for baselines?

Reviewer_z2CdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an approach to learn Unsigned Distance Field (UDF) for implicit representation of surfaces. Unlike the signed version, UDF can represent open and non-manifold surfaces, without requiring inside-outside definition. The paper starts with the premise that fitting UDF has to compromise between two related tasks: that of surface fidelity (0 field) and distance accuracy elsewhere, which is difficult to do. The authors present a variational regularizer based on the surface normals in the overall 3D space, in order to regularize the UDF field and its fidelity. In particular, the authors argue that naively optimizing for the smoothness of the normal (vector) field with Eikonal loss causes instability. This is due to the discontinuity of the actual (target) normal field on the surface and the medial axis of the surface. Therefore, the paper proposes a rescaling of the vector field by the domain average before computing the Dirichlet Energy (energy measuring the field Jacobian's norm) instead of the typical rescaling to unit norm. The regularizer is then successively annealed to obtain surface fields in the training process. The paper reports better surface reconstruction convergence with the approach compared to other approaches in single surface reconstruction.

Critical strengths
  1. The paper tackles a hard problem of UDF based implicit surface representation. Unlike SDF, UDF is extremely hard to represent and surface extraction often results in poor results due to the one sided representation's natural fidelity limit. The paper looks like a positive step in the direction with a principled approach of vector heat diffusion.

  2. The proposed variational method based on normal heat diffusion and the rescaling appears reasonable and satisfying. This is a good finding in UDF.

  3. The results appear good qualitatively for single object reconstruction.

Critical weaknesses
  1. The paper presentation seems to detach itself from the prior related works very early. In fact the Dirichlet energy is very similar to the second order energy measure in prior works but the similarities are revealed in the Appendix rather than in the main text. Also similar energy terms used by DiGS (in sdf) is mentioned somewhat reluctantly. I think they are fairly equivalent to the regularizer used here. The paper should discuss why adapting this energy due to UDF's discontinuous gradients makes it different and provide the presented solutions. Similarly, when discussing open surface reconstruction, the paper should mention alternatives to UDF such as Rella et al., Neural Vector Fields for Implicit Surface Representation and Inference, IJCV'2024 which directly uses the normal field for reconstruction that the proposed method uses inside the method.

  2. The comparisons are rather poor in metrics and analysis. The annealing and the rescaling introduces several hyperparameters whose behaviors are not analyzed carefully. This includes sensitivity to wrong hyperparameters as well as a proper guide for selecting them. Obviously one has to assume that the surface is not known beforehand when choosing the hyperparameters.

  3. Perhaps the main weakness of the method/work is apparent when considering how the method works with pre-estimated normals. In fact, a careful look at the training reveals the importance of the Dirichlet energy that relies on normal estimation which is precisely unavailable when it is most needed. The paper should describe how the normals were obtained for this training process and describe a method including this step, along with the comparisons.

  4. The surface reconstruction is not done with marching-cubes or its equivalent approaches such as MeshUDF, Guillard et al. and related works for the proposed method. Furthermore, quantitative metrics are missing in a lot of evaluations except table 2 for some samples. Furthermore, the authors use a backprop like algorithm on the field to reconstruct theirs but not the SOTA. This is a choice that favors regularized fields. Is this a requirement for the method in order to beat the SOTA or can the method perform better even with efficient meshing algorithms?

Questions for authors

Main questions are weaknesses 2 and 4, regarding the sensitivity to hyperparameters and surface reconstruction approach used. Also clarify the normal initialization.

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

Related papers

Discussion (0)

Sign in to comment.