Efficient Differentiation of Pixel Reconstruction Filters for Path-Space Differentiable Rendering

Supplemental Materials

Zihan Yu¹, Cheng Zhang^1,2, Derek Nowrouzezahrai³, Zhao Dong², and Shuang Zhao¹ ¹University of California, Irvine ²Meta Reality Labs ³McGill University

In this supplemental material, we show animated versions of a few inverse-rendering optimizations.

Left click the images below to start/pause; right click to reset the animations.

Tent filter

All results below use the tent pixel reconstruction filter.

World Map

This example involves a 3D world map.
Using 13 target images (with one shown), we jointly optimize the texture and geometry of the world map.
Estimator for the interior integral: unidirectional path tracing.
Estimator for the pixel boundary integral: N/A.

Bunny 1

This example involves a translucent bunny (without refractive interfaces).
Using 20 target images (with one shown), we optimize the shape of the bunny.
Estimator for the interior integral: unidirectional volume path tracing.
Estimator for the pixel boundary integral: N/A.

Kitty 2

This example involves a kitty lit by an area light within the Cornell box. Additionally, the light source points toward the ceiling, causing the scene to be dominated by indirect illumination.
Using 20 target images (with one shown), we optimize the shape of the kitty.
Estimator for the interior integral: bidirectional path tracing.
Estimator for the pixel boundary integral: N/A.

Tree

This example involves a tree-like object. It compares our method (utilizing antithetic sampling) with the pixel filter normalization method [Vicini et al. 2022].
Using single target image, we optimize the rotation angle of the object.
Estimator for the interior integral: unidirectional path tracing.
Estimator for the pixel boundary integral: N/A

Edge

This example involves a textured plane with two colors separated by a tilted vertical edge.
Using a single target image, we optimize the translation of the object along the vertical axis.
Estimator for the interior integral: unidirectional path tracing.
Estimator for the pixel boundary integral: N/A

Box filter

All results below use the box pixel reconstruction filter that is unsupported by previous path-space methods due to the presence of jump discontinuities.

Terrain

This example involves a terrain-like glossy object.
Using 5 target images (with one shown), we optimize the shape of the object.
Estimator for the interior integral: unidirectional path tracing.
Estimator for the pixel boundary integral: ours (described in Section 6 of the paper).

Pig 2

This example involves a pig with detailed surface geometry.
Using 20 target images (with one shown), we optimize the shape of the object.
Estimator for the interior integral: unidirectional path tracing.
Estimator for the pixel boundary integral: ours (described in Section 6 of the paper).