DANTE-W Diffuse Albedo Neural Texturing in the Wild

ECCV 2026

1Tsinghua University

arXiv Code Data
↓

TL;DR: Traditional 3D reconstruction remains a de facto in industries, this work proposes a plug-and-play diffuse albedo texturing method exploiting diffusion priors and neural rendering.

Reason beyond pixel: disentangle base colour from lighting effects

Compared to vanilla mesh texturing with baked-in lighting effects (e.g., noon-time shading and strong roof-edge shadowing on this pavilion), our method effectively disentangles a 3D-consistent diffuse albedo texture with exceptional photorealism.

Motivation overview

Issues of intrinsic decomposition diffusion model

1) 3D inconsistent

3D inconsistent diffusion result

2) inaccurate at details

Diffusion result inaccurate at details

3) incompatible with graphic pipelines

Method

Aggregate per-view outputs into a consistent neural texture

Aggregate per-view outputs into a consistent neural texture

Learning fine details by differentiating the spatial frequency of neural fields

Learning fine details by differentiating the spatial frequency of neural fields

We observe that real-world lighting effects manifest as lower frequencies compared to fine-grained albedo variations. Therefore, we represent diffuse albedo texture using a high-resolution 2D hash encoding and irradiance using a low-resolution 3D hash encoding. This explicit frequency-band discrepancy effectively facilitates the disentanglement between the two intrinsic components. We guide the low-frequency component of diffuse albedo with screen-space diffusion priors and recover fine-grained albedo details by neural rendering the raw observations.

Results

Faithful, fine-grained, 3D-consistent diffuse albedo texture

Faithful, fine-grained, 3D-consistent albedo results Challenging qualitative diffuse albedo results

Citation

@InProceedings{Wang_2024_CVPR,
  author    = {Wang, Guangyu and Zhang, Jinzhi and Wang, Fan and Huang, Ruqi and Fang, Lu},
  title     = {XScale-NVS: Cross-Scale Novel View Synthesis with Hash Featurized Manifold},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2024},
  pages     = {21029-21039}
}