generative adversarial text to image synthesis

31
Generative Adversarial Text to Image Synthesis Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran [GitHub ] [Arxiv ] Slides by Víctor Garcia [GDoc ] Computer Vision Reading Group (30/09/2016)

Upload: xavier-giro

Post on 14-Jan-2017

193 views

Category:

Data & Analytics


5 download

TRANSCRIPT

Page 2: Generative adversarial text to image synthesis

Index● Introduction ● State of the Art● Method

○ Network Architecture○ Losses

● Experiments○ Qualitative Results○ Sentence interpolation○ Style Transfer

● Conclusions

Page 3: Generative adversarial text to image synthesis

Introduction

Text → Image

GANs

Page 4: Generative adversarial text to image synthesis

Index● Introduction ● State of the Art● Method

○ Network Architecture○ Losses

● Experiments○ Qualitative Results○ Sentence interpolation○ Style Transfer

● Conclusions

Page 5: Generative adversarial text to image synthesis

GANs

Discriminator

1/0

True

World

Fake

Generator

Page 6: Generative adversarial text to image synthesis

GANs

DiscriminatorD(·)

1/0

True

World

Fake

Generator

q(x) xG(z) zx’

Page 7: Generative adversarial text to image synthesis

GANs

DiscriminatorD(·)

MAX → E[log(D(X))]

True

World

Fake

Generator

q(x) xG(z) zx’

Page 8: Generative adversarial text to image synthesis

GANs

DiscriminatorD(·)

MAX → E[log(D(X))] + E[ log(1 - D(G(Z))) ]

True

World

Fake

Generator

q(x) xG(z) zx’

Page 9: Generative adversarial text to image synthesis

GANs

DiscriminatorD(·)

MAX → E[log(D(X))] + E[ log(1 - D(G(Z))) ]

True

World

Fake

Generator

q(x) xG(z) zx’

Page 10: Generative adversarial text to image synthesis

GANs

DiscriminatorD(·)

True

World

Fake

Generator

q(x) xG(z) zx’

MIN → E[ log(1 - D(G(Z))) ]

Page 11: Generative adversarial text to image synthesis

GANs with Join DistributionsHow do we generate the image from text?

Page 12: Generative adversarial text to image synthesis

GANs with Join DistributionsHow do we generate the image from text?

Discriminator

1/0

f(x,t) f(x’,t)

Page 13: Generative adversarial text to image synthesis

GANs with Join Distributions

Discriminator

1/0

Real Image

+Text

Gen. Image

+Text

Generator +Text

Page 14: Generative adversarial text to image synthesis

GANs with Join Distributions

Discriminator

1/0

Real Image

+Text

Gen. Image

+Text

Generator +Text

Page 15: Generative adversarial text to image synthesis

Text EmbedddingIn order to represent the text in a vector...

MIN

WHERE

Page 16: Generative adversarial text to image synthesis

Text EmbedddingIn order to represent the text in a vector...

MIN

WHERE

This is the recurrent text encoder

Page 17: Generative adversarial text to image synthesis

Index● Introduction ● State of the Art● Method

○ Network Architecture○ Losses

● Experiments○ Qualitative Results○ Sentence interpolation○ Style Transfer

● Conclusions

Page 18: Generative adversarial text to image synthesis

Network Architecture

Page 19: Generative adversarial text to image synthesis

Losses - CLS

log(D(x,t)) log(1-D(G(z,t)))

True Image +

True Text

Fake Image +

True Text

Real Images match the text content?

Page 20: Generative adversarial text to image synthesis

Losses - CLS

log(D(x,t)) log(1-D(G(z,t))) log(1-D(G(zi,tk)))

True Image +

True Text

Fake Image +

True Text

True Image (i) +

True Text (j)Unmatched

Page 21: Generative adversarial text to image synthesis

Losses - INT

They train interpolating between different text embedding vector (t1~t2).

So the generator learns to fill GAPS on the data manifold.

Page 22: Generative adversarial text to image synthesis

Index● Introduction ● State of the Art● Method

○ Network Architecture○ Losses

● Experiments○ Qualitative Results○ Sentence interpolation○ Style Transfer

● Conclusions

Page 23: Generative adversarial text to image synthesis

Qualitative Results - Birds

Page 24: Generative adversarial text to image synthesis

Sentence Interpolation

Gen.

z0

+Text1

Gen.

z1

+Text3

Gen.

z0

+Text2

Gen.

z1

+Text4

Page 25: Generative adversarial text to image synthesis

Disentangling style and content

Generator.

z+

Text

If ‘text’ is describing the content? What is ‘z’ describing?

Page 26: Generative adversarial text to image synthesis

Disentangling style and content

Generator.

z+

Text

If ‘text’ is describing the content? What is ‘z’ describing?

Style → Pose, Background…, let’s extract ‘z’

Page 27: Generative adversarial text to image synthesis

Disentangling style and contentz0 z1 z2 z3 z4 z5

Page 28: Generative adversarial text to image synthesis

Qualitative Results - Flowers

Page 29: Generative adversarial text to image synthesis

Qualitative Results - MSCOCO

Page 30: Generative adversarial text to image synthesis

Conclusions

Discriminator

1/0

f(x,t) f(x’,t)

x~t

Page 31: Generative adversarial text to image synthesis