Intrinsic decomposition and editing of 3D Gaussian splats

Proceedings of the ACM Computer Graphics and Interactive Techniques

(Symposium on Interactive 3D Graphics and Games (I3D) 2026)

Alexandre Lanvin1, Jeffrey Hu1,2, Simon Lucas1, Adrien Bousseau1, George Drettakis1
1Inria, Université Côte d'Azur 2Cambridge University
Paper Code (coming soon) Group Publ. Page

Abstract

We extend intrinsic decomposition to radiance fields represented with Gaussian splatting by proposing solutions to three key aspects of such decomposition. First, we describe how to model the intrinsic decomposition as independent sets of Gaussian primitives, which allows each set to adapt to the characteristics of the layer it represents. Second, we present an optimization procedure guided by data-driven predictions to disentangle multi-view photographs of a scene into the aforementioned intrinsic sets. Finally, we provide an editing workflow where users modify the texture of planar surfaces simply by modifying the albedo of that surface in one image. Capturing this edit within the intrinsic radiance field allows re-rendering of the edited scene with plausible lighting under arbitrary viewpoints.

Method

Our method creates a 3D intrinsic decomposition of a radiance field modeled by disjoint sets of gaussians. This decomposition enables physically plausible edits by altering the albedo field while leaving the shading untouched.


3D Intrinsic Decomposition

In order to reconstruct a 3D intrinsic decomposition we first leverage DiffusionRenderer [Liang et al. 2025] to predict albedo maps from input images ➀ and train the albedo gaussians using them ➁; Then we optimize shading gaussians based on the trained albedo gaussians, input images and our diffuse image formation model ➂, finally we also optimize residual gaussians to capture any view-dependent effects.

A diffuse render is computed by taking the product of separately rendered albedo and shading gaussians. The residual layer captures view-dependent effects and it is added to the diffuse render to obtain a glossy render.

View-dependent Effects

Residual gaussians are equipped with spherical harmonics and initialized where they are the most likely to model view-dependent effects, that is where the photometric loss between our diffuse reconstruction and the input images is high.

Editing Results

BibTeX

@article{lanvin:hal-05602203,
  TITLE = {{Intrinsic decomposition and editing of 3D Gaussian splats}},
  AUTHOR = {Lanvin, Alexandre and Lucas, Simon and Hu, Jeffrey and Bousseau, Adrien and Drettakis, George},
  URL = {https://hal.science/hal-05602203},
  JOURNAL = {{Proceedings of the ACM on Computer Graphics and Interactive Techniques}},
  PUBLISHER = {{ACM}},
  YEAR = {2026},
  MONTH = May,
  DOI = {10.1145/3804495},
  KEYWORDS = {CCS Concepts: Computing methodologies $\rightarrow$ Rasterization Neural networks Reflectance modeling ; CCS Concepts: ; Computing methodologies $\rightarrow$ Rasterization ; Neural networks ; Reflectance modeling},
  PDF = {https://hal.science/hal-05602203v1/file/paper_authors_version.pdf},
  HAL_ID = {hal-05602203},
  HAL_VERSION = {v1},
}

Acknowledgments and Funding

This research was co-funded by the European Union (EU) ERC Advanced grant FUNGRAPH No 788065 and ERC Advanced Grant NERPHYS No 101141721. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the EU or the European Research Council. Neither the EU nor the granting authority can be held responsible for them. The authors are grateful to Adobe and NVIDIA for generous donations, and the OPAL infrastructure from Université Côte d’Azur. Experiments presented in this paper were carried out using the Grid'5000 testbed, supported by a scientific interest group hosted by Inria and including CNRS, RENATER and several Universities as well as other organizations (see https://www.grid5000.fr)