Heat Kernel Textures: the Geodesic Gaussians That Do Not Splat
Simone Foti, Caner Korkmaz, Stefanos Zafeiriou, Tolga Birdal
cs.CV · cs.GR · cs.LG
Quick verdict
HKTex replaces UV mapping with manifold heat kernels, achieving top SSIM/MSE at ~97KB storage, but 785ms renders leave games/AR use undemonstrated.
Summary
The paper introduces Heat Kernel Textures (HKTex), a UV-free texture representation for manifold triangle meshes that replaces texture maps with anisotropic heat kernels evaluated intrinsically on the surface. Each kernel is parametrised by a face-plus-barycentric source position, an angle and anisotropy controlling a discrete anisotropic Laplace-Beltrami operator (ALBO), a scale threshold, a sharpness, and an RGB colour. To make evaluation tractable, the authors precompute a 7x7 grid of ALBO eigendecompositions aligned via Hungarian assignment and Orthogonal Procrustes, interpolate spectra bilinearly, damp spectral ringing with a biharmonic-distance Gaussian, and restrict kernel evaluation via KNN search in a biharmonic spectral embedding. Kernel positions are optimised with Riemannian gradient descent using exponential maps and parallel transport, supported by error-driven densification and hit/energy-based pruning, inside a differentiable Mitsuba 3 pipeline. On a curated manifold-filtered subset of Objaverse (~3,000 meshes; 313 for UV fitting, 162 for multi-view), HKTex attains the best SSIM, MS-SSIM and MSE at 96.6 KB in UV fitting, and wins all quality metrics in multi-view fitting (PSNR 37.61, 78.7 KB), though downsampled ground-truth UV textures still achieve higher PSNR (48.4) and rendering takes 784.7 ms versus 25-91 ms for baselines. Main limitations: ~370 MB per-mesh spectral precomputation excluded from headline timings, albedo-only modelling, synthetic multi-view supervision with a known mesh, and evaluation restricted to clean manifold meshes.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- HKTex eliminates UV unwrapping and its associated artifacts (seams, distortion, wasted space, per-vertex UV storage) by defining textures as anisotropic heat kernels intrinsic to the mesh surface.
- A Riemannian optimisation scheme (exponential map plus parallel transport of momentum) strictly constrains kernel movement and density-control operations to the surface.
- Under strictly matched storage budgets, HKTex achieves the best SSIM, MS-SSIM and MSE in UV-texture fitting (96.6 KB) and wins all quality metrics in multi-view fitting (PSNR 37.61, 78.7 KB).
- A precomputed, alignment-corrected 7×7 grid of anisotropic-LBO eigendecompositions with biharmonic KNN pruning makes continuous kernel evaluation tractable at O(P×Ks).
- HKTex integrates into a differentiable physically based renderer (Mitsuba 3), enabling both compression of existing UV textures and inverse rendering from multi-view images.
Strengths
Genuinely intrinsic formulation with principled optimisation
Kernel positions are updated by Riemannian gradient descent on the surface: tangent-plane projection, exponential map and parallel transport of momentum via straightest geodesics (Sec. 4.3, refs [32,38]); density-control splits also displace children along the exponential map. This avoids the regulariser-based surface alignment used by MeshGS [8] and SUGAR [14].
Scalable spectral machinery that makes the idea tractable
Precomputed 7×7 half-polar grid of ALBO eigendecompositions with Hungarian assignment + Orthogonal Procrustes + sign alignment for consistent bilinear spectral interpolation (Sec. 4.1); biharmonic-distance Gaussian weighting suppresses Gibbs ringing from K=256 truncation (Eqs. 4-5); KNN in the biharmonic embedding Φ(Λ)^-1 bounds evaluation at O(P×Ks) instead of O(P×S) (Sec. 4.2).
Strong matched-storage quantitative results
Unusually honest ablations
End-to-end differentiable PBR integration
Weaknesses
Rendering cost contradicts the motivating applications
Heavy precomputation kept out of headline costs
Evaluation on a heavily curated, easy subset
Baseline comparisons inflated by forced storage mismatch
Narrow supervision scope and unquantified approximations
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
HKTex (arXiv 2609.07557, Imperial College London) proposes an intrinsic, UV-free texture representation: a mesh's appearance is stored as S optimisable anisotropic heat kernels evaluated continuously on the surface, replacing UV-mapped texture images. Central claims: (1) elimination of UV artifacts (seams, distortion, wasted atlas space, per-vertex UV coordinates and seam-duplicated vertices); (2) strict surface-constrained optimisation via geodesic tools; (3) a lower memory footprint at matched or better quality; (4) differentiable PBR integration enabling both UV-texture compression and multi-view inverse rendering. Scope is deliberately limited to albedo. This is a method paper with quantitative benchmarks on a curated Objaverse subset, positioned against neural fields, 2D Gaussian image representations and vertex colours.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Heat Kernel Textures (HKTex)?
HKTex is a texture representation for 3D meshes that stores appearance as a set of anisotropic heat kernels defined directly on the mesh surface instead of a UV-unwrapped texture image. Each kernel's position, orientation, anisotropy, size, sharpness and colour are optimised to reproduce either an existing texture or multi-view images, eliminating UV seams, distortion and UV-coordinate storage.
How much smaller are HKTex textures than UV textures?
On the paper's 313-mesh Objaverse benchmark, optimised HKTex models average 96.6 KB versus 115.8 KB for downsampled ground-truth UV textures and 473-928 KB for storage-matched neural baselines, while achieving the best SSIM, MS-SSIM and MSE among the compared methods (downsampled UV textures still win on raw PSNR, 48.4 vs 44.8).
Is HKTex code available?
Yes - the paper states the project page and code are available at circle-group.github.io/research/HeatKernelTextures. The supplementary also publishes the full hyperparameter tables used for both UV-texture and multi-view experiments.
What are the main limitations of HKTex?
Rendering is slow (784.7 ms per image vs 25-91 ms for baselines in their UV-fitting benchmark), each mesh needs a spectral precomputation averaging ~98 s and ~370 MB of cached eigenproperties that is excluded from the headline timings, only albedo is modelled, multi-view experiments are synthetic with a known mesh and lighting, and the evaluation excludes non-manifold, multi-component or very large meshes.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Figs. 1-11) are raster images and could not be inspected; qualitative claims about sharpness and artifacts rest on the quantitative tables (Tables 1, 2, 4) and the authors' descriptions.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.