1 of 16

NVIDIA RTX Ray Tracing

And Unreal Engine

2 of 16

Questions

How Did They Get This Working In Real Time?

Something Something Denoising?

3 of 16

How Did They Get This Working In Real Time?

4 Nvidia Volta GPUs ($60-100k machine)

4 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

The issue: current state of the art ray-tracers only have time to trace a few rays per pixel at 1080p/30hz. ~Doubles every few years, but counteracted by rising display resolutions. Current budget (in games, not the short) is ~1 short path per pixel.

5 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

6 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

7 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

8 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability� - End-to-end training with additional information available in the rendering setting� (depth, normals, material roughness)

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

9 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability� - End-to-end training with additional information available in the rendering setting� (depth, normals, material roughness)

Their Path Tracer:

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

10 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability� - End-to-end training with additional information available in the rendering setting� (depth, normals, material roughness)

Their Path Tracer:� - Uses “next event event estimation”

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

11 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability� - End-to-end training with additional information available in the rendering setting� (depth, normals, material roughness)

Their Path Tracer:� - Uses “next event event estimation”� - Rasterize first hits

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

12 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability� - End-to-end training with additional information available in the rendering setting� (depth, normals, material roughness)

Their Path Tracer:� - Uses “next event event estimation”� - Rasterize first hits� - Use low-discrepancy halton sequence instead of uniform sampling to reduce variance

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

13 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability� - End-to-end training with additional information available in the rendering setting� (depth, normals, material roughness)

Their Path Tracer:� - Uses “next event event estimation”� - Rasterize first hits� - Use low-discrepancy halton sequence instead of uniform sampling to reduce variance� - Maximum depth of 1, so (camera-surface-light) and (camera-surface-surface-light) paths only.

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

14 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Motivated by Deep Image Completion, but with additional features designed specifically for Monte Carlo Rendering.

Primary contributions above and beyond prior image completion work:� - RNN for temporal stability� - End-to-end training with additional information available in the rendering setting� (depth, normals, material roughness)

Their Path Tracer:� - Uses “next event event estimation”� - Rasterize first hits� - Use low-discrepancy halton sequence instead of uniform sampling to reduce variance� - Maximum depth of 1, so (camera-surface-light) and (camera-surface-surface-light) paths only.� - In order to aid training, they remove the texture from the first object hit, and then re-apply the� the texture after lighting is calculated.

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

15 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Noisy input plus normals, depth and material properties.

Reconstruction loss against smooth image rendered non-interactively.

Fully convolutional, so this can be trained on small crops (128x128)

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf

16 of 16

Interactive Reconstruction of Monte Carlo Image Sequences�using a Recurrent Denoising Autoencoder [1]

Training:� - Generate a camera fly-through of an environment.� - For each frame, generate 10 different noisy estimates as a way of collecting more data� - Render full image, but cut out 128x128 crops over seven frames � - Also can use reverse playback of the camera motion for more data augmentation� - Use L1 loss on both pixel values and image gradients� - Temporal loss to penalize flickering between frames

[1] Chaitanya et al. 2017 http://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf