acceptodds
Withdrawn from ICLR 2027

HeatAtlas: Surface Textures as Learnable Atlases of Local Heat Kernels

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Heat Kernel Textures (HKTex) represent surface appearance with learnable anisotropic heat kernels, but evaluate each kernel through a global spectral construction that requires fifty Laplace–Beltrami eigendecompositions per mesh and a resident basis of shape [50, V, 256]. We show that this evaluator is mismatched to the learned primitive: trained HKTex responses are local and lie in the short-time regime of the heat equation, where the Minakshisundaram–Pleijel expansion is the natural evaluator and depends only on local geometry. HeatAtlas replaces the global eigensystem by radius-bounded local charts from batched mesh unfolding, on which the anisotropic heat kernel becomes a closed-form geodesic Gaussian with a first-order curvature correction. We then make these charts the representation itself: each kernel is a local chart whose non-negative heat response gives a normalised overlap weight and whose appearance is a local function , instantiated with intrinsic Gabor content in a parallel-transported frame. Separating where a primitive contributes from what it contains removes the spectral mode ceiling and all global Laplacian preprocessing, and supports learnable per-kernel scale, directional high-frequency appearance, local barriers and edits, planar, disconnected, non-manifold and million-vertex inputs, and transfer across discretisations without retraining. On the 32-object structural-failure set, HeatAtlas reaches 32.29 dB against 19.16 dB for HKTex and 26.82 dB for the strongest baseline, a hash grid; across five meshes of 50k–374k vertices, initialisation is 99–206× faster and peak memory 20–45× lower at equal quality; and at matched storage, intrinsic Gabor content improves view PSNR over constant charts by up to +4.0 dB on periodic textures, and on real high-frequency assets exceeds the hash grid by +1.9 dB (LPIPS 0.044 → 0.027).

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
neural textures, heat kernel, surface appearance, geodesic Gaussians, local charts, Gabor, mesh representation, inverse rendering

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_U2BaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a novel method for learned texture representation, building on the HKTex model. It replaces the costly spectral computations with local charts and more expressive Gabor-based kernels, resulting in faster initialisation and less memory usage.

Critical strengths
  • The method is well motivated: the learned kernels are local, and spectral methods become expensive on high-resolution meshes.
  • Transfer across remeshings without retraining is useful in practice. For example, a mesh stored at several levels of detail could share a single texture. The authors could emphasise this application more.
  • The method performs well on high-frequency and periodic textures.
  • The barrier mechanism is a nice idea and could have useful applications.
Critical weaknesses
  • The main-paper tables and figures mostly use small sets: the 32-object structural-failure set, six representative scenes, a 30-object subset for rate–distortion, synthetic toy textures, and 20 objects for ablations. Results on the full 313-object UV-fitting and 162-object multi-view benchmarks only appear as tables in the appendix (Tables 8 and 11). I encourage the authors to move these into the main paper.
  • Relatedly, could the authors explain the large discrepancy between Table 8 and Table 1 of the HKTex paper? At comparable storage (~97 kB), HKTex reports 44.8 dB for HKTex and 42.5 dB for Intrinsic Neural Fields, while Table 8 reports 31.07 and 29.41 dB. The ranking also changes: the low-resolution GT texture is the strongest baseline in HKTex (48.4 dB) but the weakest here (26.10 dB).
  • It is unclear in which cases the QSlim proxy was used in table 3. A fairer comparison would also include HKTex-Simplify, since it is closer to the QSlim proxy setting. What does "O(1) per proxy face" mean exactly in table 14? The paper should provide estimates of the per-vertex cost, without the quadric approximation and as a function of mesh resolution.
  • Without the proxy, memory seems to scale roughly as O(N·V): higher vertex density means more faces inside each kernel’s support. Could the authors clarify how the charts are stored? Are A_if and b_if (or S_if) kept for every kernel-face pair, as Algorithm 1 suggests? A memory measurement as a function of |V| at fixed N, with and without the proxy, would make this clear.
Questions for authors
  • When reporting the time per step, could the authors also report the rebuild time of the unfolding as a function of |V|? For example, 5k kernels on the same mesh at several resolutions.
  • I would appreciate a more thorough analysis of the rebuild interval. Table 5 reports that rebuilding every step is 0.26 dB worse than every 10 steps, which is counter-intuitive. Could the authors explain this and provide a sweep over rebuild intervals?
  • Could the authors analyse failure cases of the unfolding compared to the Vector Heat log map, beyond the median angular error of 1.9°? For example, the error distribution near cone vertices or on high-curvature regions.

Clarifications

  • In the ablation against exact geodesics (Table 5), which quantities are replaced? MMP provides distances but not directions, so how are the chart coordinates (u, v) obtained in that variant? A comparison using a log map would be more informative.
  • The authors claim that HKTex responses are local, however there are 2 techniques used in HKTex to assess them locally: one is the multiplication with a Gaussian using BH distance, and the other one is a computational efficiency trick that evaluates only on specific KNNs.
  • Although HKTex uses a fixed diffusion time, kernel scale is controlled by a learned threshold. HKTex also fits an image on a two-triangle rectangle with only K=3 modes and still recovers fine detail (albeit with a kernel amplification exponent). This suggests the number of modes mainly affects the overall shape of the kernels rather than placing a hard limit on the finest representable detail. The “mode ceiling” argument (l. 50) should be discussed in light of this.
  • Abstract (l. 13–14): reporting “fifty eigendecompositions” and a “[50, V, 256]” basis in the abstract seems unusual, since these are tunable hyperparameters of HKTex rather than intrinsic properties.
  • l. 16–17 and l. 60: could the authors justify why the Minakshisundaram–Pleijel expansion is “the natural evaluator”?
  • l. 53: “non-manifold soups”. Do the authors mean triangle soups?
  • l. 66: the sentence “promotes the chart from an evaluation device to the representation itself” is hard to parse. Could the authors rephrase it and clarify what they mean by “evaluation device”?
  • l. 190: the statement that the spectral construction is “super-linear in |V|” seems inconsistent with the spot measurements reported right after (122 s at 2.9k vs. 237 s at 47k vertices, i.e., sub-linear), although Table 6 does show super-linear growth. Clarifying which costs dominate in each regime would resolve this.
  • The text in figures should be selectable, i.e., exported as vector graphics. Figures 2, 3, and 4, as well as Tables 1 and 2, are hard to read because the font is too small. The two stripes and two wood examples seem redundant; keeping only one of the two textures would free up space to enlarge the remaining results. Figure 4 also contains a lot of unused white-space.
Reviewer_zZheOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Observing that trained heat kernel primitives exhibit highly localized responses in the short-time regime of the heat equation, the authors propose HeatAtlas, a surface texture representation building upon HKTex. It replaces global Laplace–Beltrami eigendecompositions with radius-bounded local charts constructed through mesh unfolding, using a short-time heat kernel approximation to compute spatial blending weights and intrinsic Gabor functions to represent local texture content. This design separates the spatial support of each primitive from its appearance representation.

Experiments evaluate reconstruction quality, computational efficiency, and memory usage, as well as performance on high-frequency textures, challenging mesh structures, cross-discretization transfer, and local texture editing.

Critical strengths
  1. The paper identifies the mismatch between the global spectral evaluation used by HKTex and the localized responses of trained heat kernel primitives. Replacing global eigendecompositions with local heat kernel approximations is a reasonable design choice, and the resulting framework provides a coherent connection between local geometric computation and explicit texture representation.

  2. The experiments cover texture reconstruction quality, computational efficiency, memory usage, high-frequency patterns, challenging mesh structures, and transfer across different discretizations. The ablation studies also examine the contributions of key design choices, helping characterize the conditions under which the proposed representation is effective.

  3. HeatAtlas supports directional high-frequency texture content, local texture editing, and explicit material boundaries through edge barriers. The reported results on large meshes and structurally challenging geometries further demonstrate the practical potential of the representation beyond standard texture-fitting benchmarks.

Critical weaknesses
  1. Although the ablations examine several design choices, their benefits appear to vary across settings. A clearer characterization of the respective contributions of local evaluation, intrinsic geometry, and chart content across different tasks would strengthen the empirical justification for the overall design.
Questions for authors
  1. In Eq. (4), the distance (d_{if}) is defined as the maximum of the unfolded distance and the Euclidean distance. Could the authors clarify the geometric motivation for this choice, particularly when the two distances differ, why do we need the euclidean term, and explain how this distance relates to the local coordinates used in Eq. (5)?

  2. Appendix C briefly describes the use of a simplified proxy mesh for large inputs, with supports and candidate selection computed on the proxy while queries remain on the original mesh. Could the authors clarify how query points on the original mesh are mapped to the proxy-based local charts, and how discrepancies between the two geometries affect the accuracy of the resulting texture representation?

  3. Figure 10 demonstrates local editing using analytically initialized stripe patterns. It would be useful to clarify whether similar edits were evaluated on textures learned through optimization, where interactions between neighboring charts could introduce phase discontinuities or other artifacts. Additionally, could the authors clarify how barrier edges are specified in the material-boundary experiments?

Reviewer_1v5fOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a practical improvement over HKTex. It partitions an object's surface into many independent patches, each handling only the local geometry and stripe patterns, rather than requiring computations over the entire model as previous methods do. This substantially reduces computation time and GPU memory usage.

Critical strengths
  1. The work appears substantial, and the videos included in the supplementary material are particularly well produced.
  2. Tables 3 and 6 demonstrate substantial advantages over the baseline in terms of runtime and storage requirements.
Critical weaknesses

Overall, this paper presents a practical improvement over HKTex. However, given its incremental nature, I do not believe that its novelty and practical utility meet the standards of ICLR.

Questions for authors
  1. The original HKTex paper reports 44.8 dB across 313 objects, whereas the rerun in this paper achieves only 31.07 dB, despite having nearly the same storage footprint. What accounts for such a substantial discrepancy between the authors' reproduction and the results reported in the original paper?

  2. Your method unfolds the surface around each kernel into a small planar patch, on which weights and textures are then computed. In regions with sharp corners or high curvature, or when the same location can be reached through different unfolding paths, could this lead to inaccurate distance estimates, inconsistent stripe orientations, or texture misalignment? I would appreciate the authors' insights into the method's robustness and potential avenues for improvement.

  3. If I am not mistaken, HKTex first appeared on September 7, 2026, while the full-paper submission deadline for ICLR 2027 was September 25, 2026. Since this paper is clearly an improvement over HKTex, I am curious how the authors managed to complete this work within just 18 days, relying almost entirely on human effort rather than AI assistance.

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

Related papers

Discussion (0)

Sign in to comment.