Deep Compositional Denoising for High-quality Monte Carlo Rendering

Supplementary Materials – EGSR 2021

Xianyao Zhang, Marco Manzi, Thijs Vogels, Henrik Dahlberg, Markus Gross, Marios Papas

ETH Zürich, DisneyResearch|Studios, and Industrial Light & Magic

Pixel-based Methods - Hyperion Dataset

Comparing with KPAL baselines on a selection of scenes from our Hyperion evaluation dataset.


Shot 1

Shot 2

Shot 3

Shot 4

Shot 5

Shot 6

Shot 7

Shot 8

Shot 9

Shot 10

Shot 11

Shot 12

Shot 13

Pixel-based Methods - Mitsuba Dataset

Comparing with KPAL baselines on a selection of scenes from our Mitsuba testing dataset.


The Breakfast Room

The White Room

Salle de bain

Old vintage car

Victorian Style House

Japanese Classroom

Sponza

Kitchen 1

Kitchen 2

The Grey and White Room 1

The Grey and White Room 2

The Grey and White Room 3

Breakfast Room 1

Contemporary Bathroom

Material Test Ball

Motion Blur Ball

Teapot 1

Teapot 2

Teapot 3

House 1

Bedroom 2

Bathroom 0

Bathroom2 0

Bedroom 0

Classroom 0

Coffee 0

Breakfast Room 0

House 0

Kitchen 0

Staircase 0

Comparing with Direct-predicting Method (GAN) - Tungsten Dataset

Comparing with the GAN method [Xu et al. 2019] on a selection of scenes collected by Xu et al. (rendered with Tungsten renderer).


The Grey and White Room

The White Room

The Modern Living Room

Country Kitchen

Coffee Maker

Modern Hall

Teapot (full)

Material Test Ball

Veach Ajar

Sample-based Comparison - Mitsuba Dataset

Comparison with sample-based methods on Mitsuba testing set.


Teapot 4

The Grey and White Room 1

Sponza

Kitchen 1

Kitchen 2

Bathroom 0

Bathroom 1

Bathroom 2

Bathroom2 0

Bathroom2 1

Bathroom2 2

Bedroom 0

Bedroom 1

Bedroom 2

Classroom 0

Classroom 1

Classroom 2

Coffee 0

Coffee 1

Coffee 2

Breakfast Room 0

Breakfast Room 1

Breakfast Room 2

House 0

House 1

House 2

Kitchen 0

Kitchen 1

Kitchen 2

Staircase 0

Staircase 1

Staircase 2

Motion Blur Ball 2