GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering

Hongze Chen, Zehong Lin, Jun Zhang.

Paper | Project Page

Overview

teaser.mp4

We present GI-GS, a novel inverse rendering framework that leverages 3D Gaussian Splatting (3DGS) and deferred shading to achieve photo-realistic novel view synthesis and relighting. In our framework, we first render a G-buffer to capture the detailed geometry and material properties of the scene. Then, we perform physically-based rendering (PBR) only for direct lighting. With the G-buffer and previous rendering results, the indirect lighting can be calculated through a lightweight path tracing. Our method effectively models indirect lighting under any given lighting conditions, thereby achieving better novel view synthesis and relighting. Quantitative and qualitative results show that our GI-GS outperforms existing baselines in both rendering quality and efficiency.

Installation

1. Clone the repo

git clone https://github.com/stopaimme/GI-GS.git --recursive
cd GI-GS

2. Create the environment

conda env create --file environment.yml
conda activate GI-GS

cd submodules

git clone https://github.com/NVlabs/nvdiffrast
pip install ./nvdiffrast

pip install ./simple-knn
cd ./diff-gaussian-rasterization && python setup.py develop && cd ../..

Dataset

You can find the Mip-NeRF 360 and TensoIR-Synthetic datasets from the link provided by the paper authors. And the authors of TensoIR have provided the environment maps in this link.

Running

The training scripts heavily rely on the code provided by GS-IR. Thanks for its great work!

TensoIR-Synthetic

Take the lego case as an example.

Training

python train.py \
-m outputs/lego/ \
-s datasets/TensoIR/lego/ \
--iterations 35000 \
--eval \
--gamma \
--radius 0.8 \
--bias 0.01 \
--thick 0.05 \
--delta 0.0625 \
--step 16 \
--start 64 \
--indirect

set --gamma to enable linear_to_sRGB will cause better relighting results but worse novel view synthesis results
set --indirect to enable indirect illumination modelling

Global illumination settings

• radius → path tracing range
• bias → ensure ray hit the surface
• thick → thickness of the surface
• delta → angle interval to control the num-sample
• step → path tracing steps
• start → path tracing starting point

You can change the radius, bias, thick, delta, step, step ,start to achieve different indirect illumination and occlusion. For the precise meanings, please refer to forward.cu (line 635-910).

Evaluation (Novel View Synthesis)

python render.py \
-m outputs/lego \
-s datasets/TensoIR/lego/ \
--checkpoint outputs/lego/chkpnt35000.pth \
--eval \
--skip_train \
--pbr \
--gamma \
--indirect

Evaluation (Normal)

python normal_eval.py \
--gt_dir datasets/TensoIR/lego/ \
--output_dir outputs/lego/test/ours_None

Evaluation (Albedo)

python render.py \
-m outputs/lego \
-s datasets/TensoIR/lego/ \
--checkpoint outputs/lego/chkpnt35000.pth \
--eval \
--skip_train \
--brdf_eval

Relighting

python relight.py \
-m outputs/lego \
-s datasets/TensoIR/lego/ \
--checkpoint outputs/lego/chkpnt35000.pth \
--hdri datasets/TensoIR/Environment_Maps/high_res_envmaps_2k/bridge.hdr \
--eval \
--gamma

set --gamma to enable linear_to_sRGB will cause better relighting results but worse novel view synthesis results

Relighting Evaluation

python relight_eval.py \
--output_dir outputs/lego/test/ours_None/relight/ \
--gt_dir datasets/TensoIR/lego/

Mip-NeRF 360

Take the bicycle case as an example.

Training

python train.py \
-m outputs/bicycle \
-s datasets/nerf_real_360/bicycle/ \
--iterations 40000 \
-i images_4 \
-r 1 \
--eval \
--metallic \
--radius 0.8 \
--bias 0.01 \
--thick 0.05 \
--delta 0.0625 \
--step 16 \
--start 64 \
--degree 3 \
--indirect

-i images_4 for outdoor scenes and -i images_2 for indoor scenes
-r 1 for resolution scaling (not rescale)
-degree 3 for bicycle, flowers, stump and -degree 1 for garden, treehill and indoor scenes. We find that set SH_degree to 1 can achieve better geometry.
set --metallic choose to reconstruct metallicness
set --gamma to enable linear_to_sRGB will cause better relighting results but worse novel view synthesis results
set --indirect to enable indirect illumination modelling

Evaluation

python render.py \
-m outputs/bicycle \
-s datasets/nerf_real_360/bicycle/ \
--checkpoint outputs/bicycle/chkpnt40000.pth \
-i images_4 \
-r 1 \
--eval \
--skip_train \
--pbr \
--metallic \
--indirect

set --gamma to enable linear_to_sRGB will cause better relighting results but worse novel view synthesis results

Relighting

python relight.py \
-m outputs/bicycle \
-s datasets/nerf_real_360/bicycle/ \
--checkpoint outputs/bicycle/chkpnt40000.pth \
--hdri datasets/TensoIR/Environment_Maps/high_res_envmaps_2k/bridge.hdr \
--eval \
--gamma

set --gamma to enable linear_to_sRGB will cause better relighting results but worse novel view synthesis results

Acknowledge

• GS-IR (Provides framework)⭐
• gaussian-splatting
• nvdiffrast
• nvdiffrec

Citation

@inproceedings{chen2025gigs,
      title={GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering}, 
      author={Hongze Chen and Zehong Lin and Jun Zhang},
      booktitle={ICLR},
      year={2025},
}