Abstract: Generative models, and in particular adversarial ones, are becoming prevalent in computer vision as they enable enhancing artistic creation, inspire designers, prove usefulness in semi-supervised learning or robotics applications.
We will see how to develop the abilities of Generative Adversarial Networks (GANs) to
deviate from training examples to generate more original images of fashion designs. As a limitation of GANs is the production of raw images of low resolution, we also present solutions to produce vectorized results, and show how the developed method may be useful for image editing.
Image generative modeling for design inspiration and image editing by Camille Couprie, Research Scientist @Facebook AI
1. Generative modeling for design
inspiration and image editing
Camille Couprie
Facebook AI Research
Joint works with O. Sbai, M. Aubry, M. Riviere, A. Bordes, M. Elhoseiny, Y. LeCun
2019
1
4. 4
Research areas
Core machine learning
Theory and optimization
Reinforcement learning
Artificial general intelligence /
cognitive science
Computer Vision
Speech processing
Natural Language processing
Grounding, Interaction,
Communication
Robotics
System research
5. Why do we care about generative models
5
• Scene Understanding can be assessed by checking the ability to
generate plausible new scenes.
• Generative models are interesting if they can be used to go beyond
training data: data of higher resolution, data augmentation to help
train better classifiers, use the learned representations in other
tasks, make prediction about uncertain events, or producing original
data content that can be inspiring...
14. RANDOM
NUMBERS
Generator
0 . 3
0 . 7
0 . 1
0 . 8
AdVERSARIAL
NETWORK
Generated Image
D o t t e d
F l o r a l
g r a p h i c a l
u n i f o r m
t i l e d
s t r i p e d
A n i m a l p r i n t
With the Style
Deviation criterion
(CAN H)
15. O v e r a l L L i k a b i l i t y ( % )
6 5 7 0 7 56 0 8 0
GAN: DCGAN [Radford et al.]
CAN: GAN with "Creativity" loss [Elgammal et al.],
CAN (H) stands for the use of our "holistic" loss.
CAN
texTURE
GAN
Mask
CAN(h)
texTURE
CAN(h)
texTUREr e a l i s t i c
A p p e a r a n c e
4 8
5 3 . 5
5 9
6 4 . 5
7 0
Human Evaluation Study
16. Models with texture deviation are Most Popular
N o v e l t y0 1
L i k e a b i l i t y
6 2
6 6
7 0
7 4
7 8
judged by humans and measured as a distance to similar training images
Can
texture
Can (H)
Shape Can (H)
texture
mask can
(h)
GAN
17.
18. 1024x1024 generations on the RTW dataset
2
0
Using Morgane Riviere’s pytorch code “progressive growing of GANs”, Karras et al., ICLR’18
19. 2
1
Sbai, Couprie, Aubry, arxiv dec18
3/ Vector Image Generation
by learning parametric layer decomposition
Current deep generative models
are great but…
… are limited in resolution, and
control in generations
20. Kanan et al: Layered
GANs (LR-GANs), ICLR’17
GANIN et al. SPIRAL,
ICML’18
Related work
2
2
27. Image vectorization:
Reconstruction results on
MNIST images. Our model
learns a vectorized mask
representation of digits that
can be generated at any
resolution without
interpolation artifacts.
Vector output
3
0
32. Conclusions
3
5
Some open problems:
- Automatic metrics to evaluate generative models performances
- How to improve GANs resolution?
Torch code online :
- Pytorch GAN zoo
- Vector image generation: available soon on Othman Sbai’s github