1: \begin{abstract}

2:   Recent advances in differentiable rendering have enabled high-quality reconstruction of 3D scenes from multi-view images. 

3:   Most methods rely on simple rendering algorithms: pre-filtered direct lighting or learned representations 

4:   of irradiance. We show that a more realistic shading model, incorporating ray tracing and Monte Carlo integration, 

5:   substantially improves decomposition into shape, materials \& lighting.

6:   Unfortunately, Monte Carlo integration provides estimates with significant noise, even at large sample counts,

7:   which makes gradient-based inverse rendering very challenging.  

8:   To address this, we incorporate multiple importance sampling and denoising in a novel inverse rendering pipeline. 

9:   This improves convergence and enables gradient-based optimization at low sample counts.

10:   We present an efficient method to jointly reconstruct geometry (explicit triangle meshes), materials, and lighting,

11:   which substantially improves material and light separation compared to previous work. We argue that denoising 

12:   can become an integral part of high quality inverse rendering pipelines. 

13: \end{abstract}