The Next Leap:�How A.I. and automation might change the 3D industry
?
I almost never get asked: �"What's not going to change in the next 10 years?"
I frequently get asked: �"What's going to change in the next 10 years?"
...and that’s actually the more important of the two.�
--Jeff Bezos, Amazon CEO
Better,
Faster or
Cheaper
Any technology that makes things...
will eventually become standard.
The rising development cost of AAA games
$1B
$100M
$10M
$1M
$100K
$10K
Cost (in 2017 dollars)
1980 1985 1990 1995 2000 2005 2010 2015 2020
Gamasutra: Raph Koster’s “The cost of Games”
Assets: unreasonably expensive
Modelling: 12 hours
Texturing: 10 hours
First pass total: 22 hours
Total of 66 Hours @ $60/hr
$3900
Revisions: x2-3
Building from Turbosquid, user 3dThorium
Buildings
$3900
Buildings
$3900
Buildings
$6,000
Buildings
$6,000
Building
$3900
Building
$3900
Buildings
$3900
Waving flags
$2,250
Bannister
$2,450
Building Front
$3,800
Truck
$5,600
Car
$7,200
Car Damaged
$7,200
Car
$6,000
Traffic Cones
$3,600
Concrete
Divider:
$960
Road Texturing
$10,800
Road Texturing
$11,400
Hero Character
$49,000
Moving Sky
$4,200
Tree
$2,200
Tree
$2,200
Trash
$3,438
Big Trash
$14,400
NPC
$22,500
Car
$6,600
Billboard
$1,313
Bridge
$3,500
Tree
$2,200
Estimated Creation Cost: $200,000
Barrel $940
The problem? Static workflows
Modelling: 12 hours
Texturing: 10 hours
First pass total: 18 hours
Revisions: x2-3
Building from Turbosquid, user 3dThorium
Leap #1: Procedural Workflows
Procedural Modelling: Houdini
Procedural Lake Village by Anastasia Opera
Procedural Materials: Substance Designer
Procedural Texturing: Substance Painter
Bake
Smart Materials
Smart Masks
Now the #1 Texturing Software in the World
Art by 5518 studios
Procedural Level Design: Houdini
Finished ecosystem
Rules determine species growth
Tools for customization
Final game
Watch: Procedural World Generation of Ubisoft’s Far Cry 5
Prediction: Procedural will become standard
Modelling
Texturing
Materials
World Building
Houdini
Substance Painter
Substance Designer
Houdini
Rising game sizes
1TB by 2020
Gamasutra: Raph Koster’s “The cost of Games”
81km²
65GB in size
18,000,000,000,000,000,000 planets
7GB in size
GTA V
1,000+ artists
No Man’s Sky
13 developers
.
Static Content
Procedural
Content
(magnified x1000)
Space on hard drive
Leap #2: Machine Creep
Traditional Software
Machine Learning (example)
The key difference is its ability to learn
Input
Action
Output
Assess
Appropriate Action
Is it good?
No, try again
Input
Output
Compare with others
Yes
Machine learning thrives on
Huge Datasets &
Fast Hardware.
Which will improve over time.
Noisy image (1 sample p/pixel)
Denoised
Reference (4028 samples p/pixel)
Interactive Reconstruction of Monte Carlo Image Sequences using a Recurrent Denoising Autoencoder�Chakravarty R. Alla Chaitanya (NVIDIA), Anton Kaplanyan (NVIDIA), Christoph Schied (NVIDIA), Marco Salvi, Aaron Lefohn, Derek Nowrouzezahrai (McGill University), Timo Aila�
Render
Denoised
Denoising with Kernel Prediction and Asymmetric Loss Functions
Vogels, Thijs and Rousselle, Fabrice and McWilliams, Brian and R\"othlin, Gerhard and Harvill, Alex and Adler, David and Meyer, Mark and Nov\'ak, Jan
Rendered at 50%
Time: 2m 40s
A.I. Upsampled to 100%
Render Time: 2m 40s
Rendered at 100%
Time: 9m 56s
Created using A.I. Gigapixel
A.I Upscaled to 100%
Reference (100%)
380% faster
Mode-Adaptive Neural Networks for Quadruped Motion Control
He Zhang, Sebastian Starke, Taku Komura, Jun Saito�
Prediction: Machine Learning will become standard in software
Denoising
Uprezing
Motion capture & Animation
‘If you're not using machine learning in your software, you're doing it wrong,'
-Doug Roble (Digital Domain)
Leap #3: Machine Assisted Creativity
Experimentation takes up 50-70%
Toward Multimodal Image-to-Image Translation
J.Y. Zhu, R. Zhang, D. Pathak, �T. Darrell, A. A. Efros, O. Wang, E. Shechtman.
Input
Ground Truth
Generated Samples
Image-to-Image Demo
affinelayer.com/pixsrv/
Generated Images:
Image Inputs
Progressive Growing of GANs for Improved Quality, Stability, and Variation | NVidia Research�Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen
All generated images
Generated images from text
StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks�Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, Dimitris Metaxas�
Style Transfer
Style
Claude Monet:
Input:
A Style-Aware Content Loss for Real-time HD Style Transfer (ECCV), 2018.
Artsiom Sanakoyeu*, Dmytro Kotovenko*, Sabine Lang, Björn Ommer.
Output:
A Style-Aware Content Loss for Real-time HD Style Transfer (ECCV), 2018.
Artsiom Sanakoyeu*, Dmytro Kotovenko*, Sabine Lang, Björn Ommer.
A Style-Aware Content Loss for Real-time HD Style Transfer (ECCV), 2018.
Artsiom Sanakoyeu*, Dmytro Kotovenko*, Sabine Lang, Björn Ommer.
Fooled 39% of art historians
Prediction: Artists will use machine learning to explore new ideas
Changes in the next 5 years:
#1 Procedural workflows become standard
#3 Creative assistance
#2 Machine Learning replaces many tedious technical jobs
Are we being replaced?!
At Risk Jobs
Examples:
Safe Jobs
Examples:
Undesirable grunt tasks
Actual Art
Current
Future
Grunt Work
Creative Decisions
Programming
Box Office Movies
+5.18%
Digital TV
+13.25%
Mobile Gaming
+26.8%
PC & Console Gaming
+2.7%
Sources: Newzoo Global Games Market 2018 | CNBC “Digital TV and video industry to exceed $100 billion, study says” | Statista “Global box office revenue from 2016 to 2020”
3D Rendering and Viz expecting
25.5% growth per year till 2025
or
Double by 2022 &
Quadruple by 2025
Thank you
Special thanks to Cody Winchester for research assistance
Detects surgery problems 1000x faster
465,000 metric tons of carbon emissions saved per plane
Machine Learning Scan Analysis
Could save 1.3M lives per year
Self-Driving Cars
Generative Designed Door
Nvidia OptiX™
Rendering optimizations
#1 Denoising
Nvidia OptiX™
Currently used in Redshift, Arnold, Clarisse, Vray
Octane A.I. Denoiser
Nightmare Machine
From nightmare.mit.edu
A Style-Aware Content Loss for Real-time HD Style Transfer (ECCV), 2018.
Artsiom Sanakoyeu*, Dmytro Kotovenko*, Sabine Lang, Björn Ommer.
Admin &
Legal 40%
Programming 30%
Assets
25%
Concept
Art
5%
Typical AAA game development costs
81km²
65GB in size
18,000,000,000,000,000,000 planets
Only 7GB in size
Mordvintsev, et al., "Differentiable Image Parameterizations", Distill, 2018
Procedural Materials: Substance Designer
Procedural Materials: Substance Designer
By Eric Wiley on ArtStation
A Fully Progressive Approach to Single-Image Super-Resolution
Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung,Olga Sorkine-Hornung, Christopher Schroers
Low Noise
Low Detail Loss
Low Artifacts
Frame-to-frame consistent
The Ultimate Goal
Cycles
The Division (2016) spent an estimated*
$55M
on assets
*Roughly 40% of total
45% lighter and twice as strong.
Saves 3,180kg of fuel per partition, per year.
Denoising
Noise2Noise: Learning Image Restoration without Clean Data
Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, Timo Aila
‘If you're not using machine learning in your software, you're doing it wrong,'
-Doug Roble (Digital Domain)
Typical Asset Costs:
20-45%
My Predictions
#1 Procedural Systems become standard
#3 Artists adopt machine assisted creative exploration
#2 Machine Learning replaces many tedious tasks
2018
(Today)
2019
2020
2021
2022
2023
2024
Used to be photographed
Now 3D
Currently photographed
Soon to be 3D
@shudu.gram
Procedural Materials: Substance Designer