Heat Kernel Textures: the Geodesic Gaussians That Do Not Splat

🔥 ECCV 2026: Best Paper Award 🔥
Imperial College London

AI Generated Podcast:

Abstract

3D Gaussian Splatting has recently revolutionised novel view synthesis as well as many other 3D vision methods and applications. Drawing inspiration from this representation, we now rethink textures to overcome the main issues of UV mapping while considerably lowering their memory footprint. Heat Kernel Textures (HKTex) eliminate UV unwrapping as well as their persistent issues of wasted UV space, seams, distortions, vertex-duplication, and varying resolution. Grounded in discrete Riemannian geometry and intrinsically defined on any manifold surface discretised as a triangular mesh, HKTex uses anisotropic heat kernels as geodesic equivalents to Gaussians. Like our kernels, also the optimisation of their position and the adaptive densification strategies were redefined to operate on the surface of the object to be textureised. Our novel representation is also fully integrated with a physically based renderer and can be optimised either from existing textures or multi-view images.

HKTex

Heat Kernel Textures (HKTex) is an intrinsic, UV-free texture representation where anisotropic heat kernels act as geodesic equivalents to Gaussians that natively conform to mesh geometry. Given a mesh, a set of known cameras, and an initialisation of our heat kernels, we continuously evaluate our representation, and render it differentiably with a Mitsuba based ray tracer. Rendered pixels are compared with the corresponding pixels in the mutliview GT images. Our method can also fit existing UV textures by directly evaluating the texture at arbitrary surface locations and comparing against corresponding UV values. The loss is then backpropagated to update the shape parameters of our kernels as well as their positions which are forced to remain on the surface.

HKTex Framework Overview

Kernel Modulation

Each heat kernel is defined directly on the surface and parametrised by its source position, diffusion angle and anisotropy (controlling orientation and stretch), scale, sharpness, and RGB colour.

HKTex Kernel Modulation

Optimization & Density Control

Kernel positions are updated along the surface using Riemannian Gradient Descent with momentum (made possible only thanks to digeo). In tandem, manifold-aware controllers prune inactive kernels and dynamically densify under-reconstructed regions by cloning or splitting kernels along their principal axes on the manifold.

Results

Quantitative Spidergrams

Fitting Existing UV-Textures

Simultaneous multi-view rotation and material rendering across all benchmark shapes fitted from standard UV maps. Use the slider to rotate all objects, or click the material buttons to switch materials.

Rotation:
View 1 / 6
0° 60° 120° 180° 240° 300°
Original GT UV Textures
LR (GT) UV Textures
MLP Pos. Encoding
InstantNGP
Intrinsic Neural Fields
(GT) VTex
HR (GT) VTex
ImageGS
HKTex ours
UV texture fitting results comparison

Multi-View Inverse Rendering

Novel view synthesis and material appearance decomposition optimized directly from multi-view image observations. Use the slider to rotate all objects, or click the material buttons to switch materials.

Rotation:
View 1 / 6
0° 60° 120° 180° 240° 300°
VTex
MLP Pos. Encoding
NvDiffRec*
HKTex ours
VTex
MLP Pos. Encoding
NvDiffRec*
HKTex ours
Multi-view results comparison

BibTeX

@inproceedings{foti2026hktex,
    author    = {Foti, Simone and Korkmaz, Caner and Zafeiriou, Stefanos and Birdal, Tolga},
    title     = {Heat Kernel Textures: the Geodesic Gaussians That Do Not Splat},
    booktitle = {Computer Vision -- ECCV 2026},
    year      = {2026},
    publisher = {Springer Nature Switzerland},
    pages     = {306--323},
    doi       = {10.1007/978-3-032-37595-7_17},
}