Neural Parametric Mixtures for Path Guiding

Experimental Results

Evaluation

We compare the performance of our method (both the radiance-based version and the product sampling version) with Practical Path Guiding [Müller et al., 2017] and Variance-aware Path Guiding [Rath et al., 2020], as well as unidirectional path tracing (with BSDF importance sampling). Tthe PPG-based guiding techniques is re-implemented on our custom GPU renderer. Note that the pixel weighting scheme (e.g., the inverse-variance weighting proposed by Müller et al.) and the selection probability learning is disabled for all the guiding methods.

In the experiments, we focus on equal-sample-count comparisons, since we implement all the methods on our GPU renderer, making their runtimes strongly depend on hardware and specific implementation/configurations. All the images are rendered for 750 spp, at the resolution of 1280x720. We show the comparisons on 10 scenes. The method names "Ours (radiance)" and "Ours (full)", denote the NPM-radiance and NPM-product in the main paper. The detailed statistics, including the metrics, absolute timing, are available in Tab. 2 within the main paper.

Test Scenes

Veach Door

Living Room

Bathroom

Bedroom

Staircase

Salle de Bain

Breakfast Room

White Room

Pink Room

Veach Egg