Practical Product Path Guiding Using Linearly Transformed Cosines

Stavros Diolatzis1, Adrien Grusson2, Wenzel Jakob3, Derek Nowrouzezahrai2, George Drettakis1

1 Université Côté d'Azur, Inria, 2 McGill University, 3 EPFL

Comparison between Ours and [HEV*16]

We do not use Russian roulette and set the BSDF sampling probability to 0.25 for both techniques. The training time for the GMMs are shown below:

Scene Training time
Necklace 59 sec
Glossy Kitchen 172 sec
Living room 143 sec
Attic 177 sec

Living room (5 min)

Glossy Kitchen (10 min)

Necklace (5 min)

Attic (10 min)

Comparison between Ours and [MGN17]

Living room (2 min)

Glossy Kitchen (5 min)

Necklace (3 min)

Pink Kitchen (3 min)

Attic (3 min)

Bathroom (3 min)

Comparison between BSDF sampling and LTC based BSDF sampling

We demonstrate same sample results of the two sampling methods. For the LTC based BSDF sampling we do the same hierarchical sampling scheme but ignoring the incoming radiance and ensuring we reach a certain depth.

Living room (150 spp)

Glossy Kitchen (150 spp)

Necklace (150 spp)

Pink Kitchen (150 spp)

Attic (150 spp)

Bathroom (150 spp)