atlas

Research map / 3D vision / Gaussian splatting

IRGS: Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing

CVPR 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 17 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes IRGS, an inverse rendering framework that applies the full rendering equation to 2D Gaussian splatting using a differentiable 2D Gaussian ray tracer to compute visibility and indirect radiance on the fly.

Problem

Inverse rendering with Gaussian splatting usually simplifies the rendering equation or uses learnable parameters for incident light, which fails to capture inter-reflection and yields inaccurate material and lighting estimates.

Approach

Ray tracing directly over 2D Gaussian primitives (with well-defined ray-splat intersections) lets a pretrained 2DGS model be used as a differentiable ray tracer, so the full rendering equation with Monte Carlo sampling and indirect light can be evaluated and optimized, plus a split-sum-based indirect query for relighting.

Key results

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use IRGS instead of R3DGinverse rendering on Synthetic4Relight · Synthetic4RelightRelighting PSNR: 31.00 -> 34.90Helped
use IRGS instead of R3DGinverse rendering on Synthetic4Relight · Synthetic4RelightNVS PSNR: 36.80 -> 35.48Hurt
use IRGS instead of R3DGinverse rendering on TensoIR · TensoIRRelighting PSNR: 27.367 -> 29.907Helped
use IRGS instead of R3DGinverse rendering on TensoIR · TensoIRNVS PSNR: 37.343 -> 35.876Hurt
add indirect-light gradients through ray tracing (ablation)IRGS · Synthetic4Relight (four-scene average)Relighting PSNR: 34.22 -> 34.68Helped
include indirect light during training (ablation)IRGS · Synthetic4Relight (four-scene average)Albedo PSNR: 30.10 -> 30.81Helped
include indirect light during training (ablation)IRGS · Synthetic4Relight (four-scene average)Relighting PSNR: 33.93 -> 34.68Helped
query indirect radiance during relighting (ablation)IRGS · Synthetic4Relight (four-scene average)Relighting PSNR: 33.84 -> 34.68Helped
increase sampled rays from Nr = 16 to Nr = 256 (ablation)IRGS · Synthetic4Relight (four-scene average)Relighting PSNR: 29.46 -> 34.68Helped
increase sampled rays from Nr = 64 to Nr = 256 (ablation)IRGS · Synthetic4Relight (four-scene average)NVS PSNR: 34.98 -> 35.48Helped

Surprising findings

Limitations

Experimental setup

Datasets: Synthetic4Relight TensoIR Stanford-ORB

Benchmarks: Synthetic4Relight TensoIR Stanford-ORB

Metrics: PSNR SSIM LPIPS MAE MSE

Baselines: NeRFactor InvRender TensoIR GS-IR R3DG

Builds on: 2D Gaussian splatting 3DGRT R3DG GS-IR Nvdiffrec

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy