AEGIR: Modeling Area Emitters for Indoor Inverse Rendering Using Gaussian Splatting

Mohamed Shawky Sabae, Philipp Langsteiner, Jan-Niklas Dihlmann, Hendrik Lensch

University of Tübingen

arXiv Preprint, 2026

Paper Code Coming soon
AEGIR models indoor illumination with explicit local area emitters, improving illumination estimation, intrinsic decomposition, and controllable scene editing.

Abstract

Inverse rendering is ill-posed because illumination, geometry, and material appearance are entangled in the input images. Existing relightable Gaussian Splatting methods often approximate illumination with point lights, environment maps, or implicit light fields, which makes finite-area lights and soft local shadows difficult to recover.

AEGIR introduces explicit local area emitters into a relightable Gaussian Splatting representation. By jointly optimizing emitters, geometry, and PBR materials with differentiable deferred rendering and visibility tracing, AEGIR improves lighting reconstruction and enables controllable scene editing.

Method Overview

AEGIR represents indoor light sources as localized anisotropic 3D primitives with learnable position, orientation, color, scale, and angular emission. This representation lets lighting explain shadows and highlights while reducing their leakage into albedo.

Area Emitters

Each emitter uses anisotropic scale and a Super-Gaussian angular falloff, allowing one primitive to vary from broad omnidirectional emission to more directional fixtures.

Deferred Rendering

A deferred G-buffer stores albedo, roughness, metallic, and surface normals. A microfacet BRDF evaluates illumination with emitter sampling and visibility tracing.

Joint Optimization

Photometric losses, diffusion material priors, edge-aware smoothness, light regularization, and a final emitter-only refinement stabilize the decomposition.

Optimization Pipeline

AEGIR optimization framework
Parameterized area emitters are initialized from multi-view observations and jointly optimized with Gaussian geometry and PBR materials.

Area Emitter Formulation

AEGIR models each light as a learnable ellipsoidal emitter. Its spatial scale controls the emitter shape, the angular spread controls how broad or directional the light is, and the Super-Gaussian falloff controls whether the illumination decays softly or sharply.

Emitter Parameterization

Effect of AEGIR area emitter parameters
Emitter scale, angular spread, and falloff.
AEGIR area emitter sampling and visibility tracing
Emitter sampling and visibility tracing.
Area emitter parameterization and sampling strategy used for soft local illumination.

Lighting Estimation Results

AEGIR evaluates lighting in a controlled Mitsuba setup by fixing ground-truth geometry and materials and re-rendering each scene with the recovered lights. This isolates illumination quality, showing that explicit area emitters better match the reference lighting and recover smoother indoor illumination than point-light or environment-map baselines.

Lighting Reconstruction

Reference AEGIR (ours) GS-ID IRGS Indoor
Ground-truth indoor lighting reference
AEGIR indoor lighting estimation result AEGIR indoor direct illumination inset
GS-ID indoor lighting estimation result GS-ID indoor direct illumination inset
IRGS indoor lighting estimation result IRGS indoor direct illumination inset
Environment
Ground-truth environment lighting reference
AEGIR environment lighting estimation result AEGIR environment direct illumination inset
GS-ID environment lighting estimation result GS-ID environment direct illumination inset
IRGS environment lighting estimation result IRGS environment direct illumination inset
Lighting reconstruction under indoor and environment illumination. Insets show recovered direct illumination maps for each estimated lighting representation.
Bedroom
Livingroom
Direct illumination visualizations for the Bedroom and Livingroom scenes, showing that AEGIR recovers spatially consistent illumination maps across different viewpoints.
Method Indoor Lighting Environment Lighting
PSNR SSIM LPIPS PSNR SSIM LPIPS
GS-ID 28.35 0.92 0.11 26.35 0.86 0.15
IRGS 11.26 0.49 0.25 28.98 0.88 0.14
AEGIR 31.42 0.94 0.06 30.22 0.92 0.07

Intrinsic Decomposition and View Synthesis Results

Across real-world benchmarks, AEGIR delivers the best novel-view synthesis scores on ScanNet++, Replica, and FIPT-Real. On synthetic scenes, its improved albedo recovery indicates stronger material-illumination disentanglement, enabled by explicit area emitters and the proposed optimization strategy.

AEGIR albedo versus render PSNR plot
AEGIR improves material accuracy while preserving novel-view synthesis quality.

Intrinsic Decomposition Comparisons

FIPT Real Classroom

Property Reference AEGIR (ours) GS-ID IRGS IRIS NeILF++
Render
Classroom ground-truth RGB reference
Classroom AEGIR render
Classroom GS-ID render
Classroom IRGS render
Classroom IRIS render
Classroom NeILF++ render
Albedo
Classroom AEGIR albedo
Classroom GS-ID albedo
Classroom IRGS albedo
Classroom IRIS albedo
Classroom NeILF++ albedo
Roughness
Classroom AEGIR roughness
Classroom GS-ID roughness
Classroom IRGS roughness
Classroom IRIS roughness
Classroom NeILF++ roughness
Metallic
Classroom AEGIR metallic
Classroom GS-ID metallic
 N/A
Classroom IRIS metallic
Classroom NeILF++ metallic

FIPT Real Conferenceroom

Property Reference AEGIR (ours) GS-ID IRGS IRIS NeILF++
Render
Conference Room ground-truth RGB reference
Conference Room AEGIR render
Conference Room GS-ID render
Conference Room IRGS render
Conference Room IRIS render
Conference Room NeILF++ render
Albedo
Conference Room AEGIR albedo
Conference Room GS-ID albedo
Conference Room IRGS albedo
Conference Room IRIS albedo
Conference Room NeILF++ albedo
Roughness
Conference Room AEGIR roughness
Conference Room GS-ID roughness
Conference Room IRGS roughness
Conference Room IRIS roughness
Conference Room NeILF++ roughness
Metallic
Conference Room AEGIR metallic
Conference Room GS-ID metallic
 N/A
Conference Room IRIS metallic
Conference Room NeILF++ metallic

FIPT Synthetic Livingroom

Property Reference AEGIR (ours) GS-ID IRGS IRIS NeILF++
Render
Livingroom ground-truth RGB reference
Livingroom AEGIR render
Livingroom GS-ID render
Livingroom IRGS render
Livingroom IRIS render
Livingroom NeILF++ render
Albedo
Livingroom AEGIR albedo
Livingroom GS-ID albedo
Livingroom IRGS albedo
Livingroom IRIS albedo
Livingroom NeILF++ albedo
Roughness
Livingroom AEGIR roughness
Livingroom GS-ID roughness
Livingroom IRGS roughness
Livingroom IRIS roughness
Livingroom NeILF++ roughness
Metallic
Livingroom AEGIR metallic
Livingroom GS-ID metallic
 N/A
Livingroom IRIS metallic
Livingroom NeILF++ metallic

Replica Room 0

Property Reference AEGIR (ours) GS-ID IRGS IRIS NeILF++
Render
Room 0 ground-truth RGB reference
Room 0 AEGIR render
Room 0 GS-ID render
Room 0 IRGS render
Room 0 IRIS render
Room 0 NeILF++ render
Albedo
Room 0 AEGIR albedo
Room 0 GS-ID albedo
Room 0 IRGS albedo
Room 0 IRIS albedo
Room 0 NeILF++ albedo
Roughness
Room 0 AEGIR roughness
Room 0 GS-ID roughness
Room 0 IRGS roughness
Room 0 IRIS roughness
Room 0 NeILF++ roughness
Metallic
Room 0 AEGIR metallic
Room 0 GS-ID metallic
 N/A
Room 0 IRIS metallic
Room 0 NeILF++ metallic

Replica Office 4

Property Reference AEGIR (ours) GS-ID IRGS IRIS NeILF++
Render
Office 4 ground-truth RGB reference
Office 4 AEGIR render
Office 4 GS-ID render
Office 4 IRGS render
Office 4 IRIS render
Office 4 NeILF++ render
Albedo
Office 4 AEGIR albedo
Office 4 GS-ID albedo
Office 4 IRGS albedo
Office 4 IRIS albedo
Office 4 NeILF++ albedo
Roughness
Office 4 AEGIR roughness
Office 4 GS-ID roughness
Office 4 IRGS roughness
Office 4 IRIS roughness
Office 4 NeILF++ roughness
Metallic
Office 4 AEGIR metallic
Office 4 GS-ID metallic
 N/A
Office 4 IRIS metallic
Office 4 NeILF++ metallic
Qualitative intrinsic decomposition and novel-view synthesis on FIPT Real Classroom, FIPT Real Conferenceroom, FIPT Synthetic Livingroom, Replica Room 0, and Replica Office 4. Rows show render, albedo, roughness, and metallic outputs. N/A marks properties not produced by a method.

Scene Editing Applications

Because illumination is represented as editable geometric emitters, AEGIR supports controlled relighting and plausible virtual object insertion under local indoor lighting.

Render, Relight, and Insertion Demos

AEGIR bedroom render
Bedroom render
AEGIR bedroom relighting
Bedroom relighting
AEGIR bedroom virtual object insertion
Bedroom insertion
AEGIR Livingroom render
Livingroom render
AEGIR Livingroom relighting
Livingroom relighting
AEGIR Livingroom object insertion
Livingroom insertion
AEGIR Office 4 render
Office 4 render
AEGIR Office 4 relighting
Office 4 relighting
AEGIR Office 4 virtual object insertion
Office 4 insertion
Scene editing applications across Bedroom, Livingroom, and Office 4.

BibTeX

@misc{sabae2026aegirmodelingareaemitters,
      title={AEGIR: Modeling Area Emitters for Indoor Inverse Rendering using Gaussian Splatting}, 
      author={Mohamed Shawky Sabae and Philipp Langsteiner and Jan-Niklas Dihlmann and Hendrik Lensch},
      year={2026},
      eprint={2606.28635},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.28635}, 
}

Acknowledgments

The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Mohamed Shawky Sabae, Philipp Langsteiner, and Jan-Niklas Dihlmann.