Copper Hairball

The glossy Copper Hairball was deliberately placed inside of a glass cube to create pathologically difficult light transport under directional illumination by the sun. Unbiased flavors of path tracing - including bidrectional path tracing - can not sample the product of the BSDF and incident illumination well. While radiance-based path guiding offers a significant improvement over unidirectional path tracing and primary-sample-space path-sampling techniques, product-based path guiding is required to fully capture the difficult light transport on the copper hairball itself. We demonstrate that neural networks can leverage additional features, such as the surface normal, to capture the product even on high-frequency geometry such as the hairball (NPG-Product), outperforming competing methods at an equal sample count.
Use mouse wheel to zoom in/out, click and drag to pan. Press keys [1], [2], ... to switch between individual images. Press [?] to see more keybindings.

Charts

Lower values are better (for SSIM, we plot 1-SSIM).
MAPE
SMAPE
L1
MRSE
L2
1-SSIM

Metrics

MetricPath TracingPPGGMMPSSPSNPS KLNPG-Radiance KLNPG-Product KL scalarNPG-Product KLNPG-Product χ²
Mega Samples472472472472472472472472472
Render Time2.0m1.8m13m2.0m6.9m4.9m15m16m17m
MAPE0.4680.1280.1560.3140.1960.1420.09160.08550.138
SMAPE0.4770.1180.1370.3000.1830.1300.08440.07790.133
L10.05780.01180.01430.03210.01740.01298.83e−38.08e−30.0131
MRSE0.2890.07250.1160.1790.1190.08640.04640.04420.0568
L28.39e−31.36e−32.17e−33.87e−32.36e−31.62e−39.64e−48.90e−41.14e−3
1-SSIM0.7260.2100.2330.6150.3260.2340.1520.1420.257