ADELE: Adaptive End-to-End Mesh Optimization

Preprint · 2026

Johannes Weidenfeller1,  Shaofei Wang2,  Philipp Fürnstahl3,  Siyu Tang1

1ETH Zurich  ·  2Beijing Institute for General Artificial Intelligence  ·  3University of Zurich

Paper arXiv Code coming soon
Teaser figure — save it as  static/images/teaser.jpg 
ADELE reconstructs meshes directly. Instead of extracting a surface from a radiance field as a post-process, ADELE optimizes an adaptive Delaunay-triangulated mesh end-to-end, yielding detailed geometry with compact triangle budgets — in object-centric and unbounded scenes alike.

Abstract

Meshes remain the most practical representation for geometry reasoning and integration into graphics pipelines, yet existing reconstruction methods struggle to produce high-quality results. Most state-of-the-art approaches leverage radiance fields (NeRF/3DGS) and treat mesh extraction as a post-processing step, which often leads to either oversmoothed surfaces or excessive triangle counts. Existing end-to-end mesh optimization methods alleviate some of these issues but suffer from poor scalability and an unstable optimization process. The former limits the application scope to small-scale, object-centric scenes, while the latter leads to overall inferior reconstruction quality compared to radiance-field-based approaches.

In this paper, we introduce an adaptive, end-to-end mesh optimization framework and a novel mesh rendering technique to address these challenges, respectively. Our representation combines a Delaunay-triangulated grid with a multi-resolution hash encoding, initialized from VGGT depth maps and refined through point pruning and insertion. A depth-offset rendering formulation improves optimization stability and reduces geometric artifacts. Our method significantly outperforms existing end-to-end mesh optimization approaches on DTU and Tanks and Temples while remaining competitive with radiance-field-based methods. We also show that, for the first time, native mesh representations can effectively reconstruct detailed geometry in unbounded scenes from the Tanks and Temples dataset.

Method

Pipeline figure — save it as  static/images/method.jpg 
A Delaunay-triangulated grid carries a multi-resolution hash encoding, is initialized from VGGT depth maps, and is refined by point pruning and insertion during optimization. Depth-offset rendering stabilizes the gradients that drive the surface.

Results

Qualitative comparisons —  static/images/results.jpg 
Comparisons on DTU and Tanks and Temples against end-to-end mesh optimization and radiance-field-based baselines.

BibTeX

@article{weidenfeller2026adele,
  title   = {ADELE: Adaptive End-to-End Mesh Optimization},
  author  = {Weidenfeller, Johannes and Wang, Shaofei and
             F{\"u}rnstahl, Philipp and Tang, Siyu},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}