Top Banner
GAN Architectures and Conditional GANs Prof. Leal-Taixé and Prof. Niessner 1
75

GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Aug 22, 2020

Download

Documents

dariahiddleston
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

GAN Architectures and Conditional GANs

Prof. Leal-Taixé and Prof. Niessner 1

Page 2: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

GAN Architectures

Prof. Leal-Taixé and Prof. Niessner 2

Page 3: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Multiscale GANs

Credit: Li/Karpathy/JohnsonProf. Leal-Taixé and Prof. Niessner 3

Page 4: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Multiscale GANs

Credit: Li/Karpathy/JohnsonProf. Leal-Taixé and Prof. Niessner 4

Page 5: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Progressive Growing GANs

https://github.com/tkarras/progressive_growing_of_gans [Karras et al. 17]

https://github.com/tkarras/progressive_growing_of_gans
Page 6: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

2x

8×8

8×8

2x

16×16

16×16

2x

32×3232×32

4×4

4×4

G

Nearest-neighbor upsampling

3×3 convolution

Replicated block

Page 7: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

2x

16×16

16×16

2x

32×3232×32

2x

8×8

8×8

G

toRGB

1×1 convolution

4×4

4×4

Page 8: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

toRGB

toRGB

2x

16×16

16×16

2x

32×3232×32

4×4

4×4

G

2x

8×8

8×8

Page 9: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

toRGB

toRGB

2x

16×16

16×16

2x

32×3232×32

4×4

4×4

G

2x

8×8

8×8+

Linear crossfade

Page 10: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

2x

16×16

16×16

2x

32×3232×32

2x

8×8

8×8

4×4

4×4

G

toRGB

Page 11: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

2x

16×16

16×16

2x

32×3232×32

2x

8×8

8×8

4×4

4×432×3232×32

0.5x

16×16

16×160.5x

8×8

4×4

fromRGB

8×80.5x

4×4

G D

toRGB

Page 12: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]
Page 13: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Progressive Growing GANs

https://github.com/tkarras/progressive_growing_of_gans [Karras et al. 17]

https://github.com/tkarras/progressive_growing_of_gans
Page 14: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Lots of GAN Variations• Hundreds of GAN papers in the last two years

– > Mostly with different losses– > Extremely hard to train and evaluate

Page 15: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Conditional Generative Adversarial

Networks (cGANs)Prof. Leal-Taixé and Prof. Niessner 24

Page 16: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Conditional GANs (cGANs)• Gain control of output

• Modeling (e.g., sketch-based modeling, etc.)– Add semantic meaning to latent space manifold

• Domain transfer– Labels on A -> transfer to B, train network on ‘B’, test on B– More later

Prof. Leal-Taixé and Prof. Niessner 25

Page 17: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

GAN Manifold

Prof. Leal-Taixé and Prof. Niessner 26[Radford et al. 15]

Train Data Sampled Data -> G(z)

Page 18: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

GAN Manifold

Prof. Leal-Taixé and Prof. Niessner 27[Bojanowski et al 17]

Page 19: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

GAN Manifold

Prof. Leal-Taixé and Prof. Niessner 28

Page 20: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

GAN Manifold

Prof. Leal-Taixé and Prof. Niessner 29[Radford et al. 15]

𝐺(𝑧0) 𝐺(𝑧1)Linear interpolation in z space: 𝐺(𝑧0 + 𝑡 ⋅ 𝑧1 − 𝑧0 )

Page 21: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Conditional GANs (cGANs)

Prof. Leal-Taixé and Prof. Niessner 30Slide credit Zhu

Page 22: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Overview

original photo

projection on manifold

Project Edit Transfer

transition between the original and edited projection

different degree of image manipulation

Editing UI

Slide credit Zhu / [Zhu et al. 16]

Page 23: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Overview

original photo

projection on manifold

Project Edit Transfer

transition between the original and edited projection

different degree of image manipulation

Editing UI

Slide credit Zhu / [Zhu et al. 16]

Page 24: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

0.196 0.238 0.332

Optimization

iGANs: Projecting an Image onto the Manifold

Input: real image 𝑥𝑅

Output: latent vector z

Generative model 𝐺(𝑧)Reconstruction loss 𝐿

Slide credit Zhu / [Zhu et al. 16]

Page 25: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

0.196 0.238 0.332

Optimization

iGANs: Projecting an Image onto the Manifold

Inverting Network z = 𝑃 𝑥

0.218 0.242 0.336Auto-encoderwith a fixed decoder G

Input: real image 𝑥𝑅

Output: latent vector z

Slide credit Zhu / [Zhu et al. 16]

Page 26: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

0.196 0.238 0.332

Optimization

iGANs: Projecting an Image onto the Manifold

Inverting Network z = 𝑃 𝑥

0.218 0.242 0.336

0.153 0.167

Hybrid MethodUse the network as initialization

for the optimization problem0.268

Input: real image 𝑥𝑅

Output: latent vector z

Slide credit Zhu / [Zhu et al. 16]

Page 27: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Overview

original photo

projection on manifold

Project Edit Transfer

transition between the original and edited projection

different degree of image manipulation

Editing UI

Slide credit Zhu / [Zhu et al. 16]

Page 28: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Manipulating the Latent Vector

Objective:

𝐺(𝑧)

Guidance 𝑣𝑔

𝑧0

user guidance imageconstraint violation loss 𝐿𝑔

Slide credit Zhu / [Zhu et al. 16]

Page 29: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Overview

original photo

projection on manifold

Project Edit Transfer

transition between the original and edited projection

different degree of image manipulation

Editing UI

Slide credit Zhu / [Zhu et al. 16]

Page 30: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Edit Transfer

𝐺(𝑧1)𝐺(𝑧0)

Input

Motion (u, v)+ Color (𝑨𝟑×𝟒): estimate per-pixel geometric and color variation

Linear Interpolation in 𝑧 space

Slide credit Zhu / [Zhu et al. 16]

Page 31: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Edit Transfer

𝐺(𝑧1)𝐺(𝑧0)

Input

Linear Interpolation in 𝑧 space

Motion (u, v)+ Color (𝑨𝟑×𝟒): estimate per-pixel geometric and color variation

Slide credit Zhu / [Zhu et al. 16]

Page 32: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

iGANs: Edit Transfer

Result

𝐺(𝑧1)𝐺(𝑧0)

Input

Motion (u, v)+ Color (𝑨𝟑×𝟒): estimate per-pixel geometric and color variation

Linear Interpolation in 𝑧 space

Page 33: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

cGANs: Interactive GANs

https://github.com/junyanz/iGAN [Zhu et al. 16.]

Interactive GANs: projection to GAN embedding

Prof. Leal-Taixé and Prof. Niessner 42

https://github.com/junyanz/iGAN
Page 34: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

cGANs: Interactive GANs

Prof. Leal-Taixé and Prof. Niessner 43https://github.com/junyanz/iGAN [Zhu et al. 16.]

https://github.com/junyanz/iGAN
Page 35: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

cGANs: Interactive GANs

Prof. Leal-Taixé and Prof. Niessner 44https://github.com/junyanz/iGAN [Zhu et al. 16.]

https://github.com/junyanz/iGAN
Page 36: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Mapping in Latent Space is Difficult!• Semantics are missing• In most cases, no labels available • Ideally, need some unsupervised disentangled rep.

Prof. Leal-Taixé and Prof. Niessner 45InfoGAN [Chen et al. 16]

Page 37: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Paired vs Unpaired Setting

Prof. Leal-Taixé and Prof. Niessner 46

Page 38: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

pix2pix: Image-to-Image Translation

slides credit: Isola / Zhu

Page 39: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

real or fake?

[Goodfellow et al. 2014]

Discriminator

z G(z)

D

Generator

G

slides credit: Isola / Zhu

min𝐺

max𝐷

𝔼𝑧,𝑥 log𝐷(𝐺 𝑧 ) + log(1 − 𝐷 𝑥 )

Page 40: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

real or fake?

Discriminator

x G(x)

D

Generator

G

min𝐺

max𝐷

𝔼𝑥,𝑦 log 𝐷(𝐺 𝑥 ) + log(1 − 𝐷 𝑦 )

slides credit: Isola / Zhu

Page 41: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Real!

Discriminator

x G(x)

D

Generator

G

min𝐺

max𝐷

𝔼𝑥,𝑦 log 𝐷(𝐺 𝑥 ) + log(1 − 𝐷 𝑦 )

slides credit: Isola / Zhu

Page 42: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Discriminator

x G(x)

D

Generator

G Real too!

min𝐺

max𝐷

𝔼𝑥,𝑦 log 𝐷(𝐺 𝑥 ) + log(1 − 𝐷 𝑦 )

slides credit: Isola / Zhu

Page 43: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

min𝐺

max𝐷

𝔼𝑥,𝑦 log 𝐷(𝑥, 𝐺 𝑥 ) + log(1 − 𝐷 𝑥, 𝑦 )

real or fake pair ?

x G(x)

G

D

match joint distribution p G x , y ∼ p(x, y)fake pair real pair

slides credit: Isola / Zhu

Page 44: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

pix2pix

53

Page 45: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Edges → Images

Input Output Input Output Input Output

Edges from [Xie & Tu, 2015]slides credit: Isola / Zhu

Page 46: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

pix2pix: Paired Setting

• Great when we have ‘free’ training data

• Often called self-supervised

• Think about these settings

55

Page 47: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Sketches → Images

Input Output Input Output Input Output

Trained on Edges → Images

Data from [Eitz, Hays, Alexa, 2012]slides credit: Isola / Zhu

Page 48: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

#edges2cats [Christopher Hesse]

Ivy Tasi @ivymyt

@gods_tail

@matthematician

https://affinelayer.com/pixsrv/

Vitaly Vidmirov @vvid

slides credit: Isola / Zhu

Page 49: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Input Output Groundtruth

Data from[maps.google.com]

slides credit: Isola / Zhu

http://maps.google.com/
Page 50: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

BW → Color

Input Output Input Output Input Output

Data from [Russakovsky et al. 2015]slides credit: Isola / Zhu

Page 51: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Ideas behind Pix2Pix• 𝐿 = 𝐿𝐺𝐴𝑁 + 𝜆𝐿1 (makes it more constraint)

• Unet / skip connections for preserving structure

• Noise only through dropout– cGANs tend to learn to ignore the random vector z– Still want probabilistic model

Prof. Leal-Taixé and Prof. Niessner 60

Page 52: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Ideas behind Pix2Pix• L1 or L2 loss for low frequency details

• GAN discriminator for high frequency details

-> PatchGAN– GAN discriminator applied only to local patches– It’s fully-convolutional; i.e., can run on arbitrary image

sizes

Prof. Leal-Taixé and Prof. Niessner 61

Page 53: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Pix2PixHD• Expand the pix2pix idea to multi-scale

• Coarse-to-fine generator + discriminator

• G’s and D’s are the same but since they operate on different resolutions, they have effectively a larger receptive field

Prof. Leal-Taixé and Prof. Niessner 62[Wang et al. 18]

Page 54: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Pix2PixHD

Prof. Leal-Taixé and Prof. Niessner 63[Wang et al. 18]

Page 55: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Pix2PixHD• Use of multi-scale discriminators

• min𝐺

max𝐷1,𝐷2,𝐷3

σ𝑘=1,2,3𝐿𝐺𝐴𝑁 (𝐺, 𝐷𝑘)

• Can make various combinations of stacking discriminator and generator– E.g., have a single G and downsample generated and real

images – or have intermediate real images (cf. ProGAN)

Prof. Leal-Taixé and Prof. Niessner 64[Wang et al. 18]

Page 56: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Pix2PixHD

Prof. Leal-Taixé and Prof. Niessner 65[Wang et al. 18]

Page 57: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Pix2PixHD

Prof. Leal-Taixé and Prof. Niessner 66[Wang et al. 18]

Page 58: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Pix2PixHD (Interactive Results)

Prof. Leal-Taixé and Prof. Niessner 67[Wang et al. 18]

Page 59: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

- Expensive to collect pairs.- Impossible in many scenarios.

Label ↔ photo: per-pixel labeling

…

Paired

Horse ↔ zebra: how to get zebras?

slides credit: Isola / Zhu

Page 60: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

………

Paired Unpaired

slides credit: Isola / Zhu

Page 61: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

x G(x)

Generator

G

D

No input-output pairs!

slides credit: Isola / Zhu

Page 62: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Discriminator

x G(x)

D

Generator

G Real!

slides credit: Isola / Zhu

Page 63: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Discriminator

x G(x)

D

Generator

G Real too!

GANs doesn’t force output to correspond to input

slides credit: Isola / Zhu

Page 64: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

mode collapse!

slides credit: Isola / Zhu

Page 65: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

⋯

Cycle-Consistent Adversarial Networks

⋯

[Zhu*, Park*, Isola, and Efros, ICCV 2017]slides credit: Isola / Zhu

Page 66: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Cycle-Consistent Adversarial Networks

[Mark Twain, 1903]

⋯ ⋯

[Zhu*, Park*, Isola, and Efros, ICCV 2017]slides credit: Isola / Zhu

Page 67: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Cycle Consistency Loss

G(x) F(G x )x

F G x − x1

[Zhu*, Park*, Isola, and Efros, ICCV 2017]

DY(G x )

Reconstructionerror

slides credit: Isola / Zhu

Page 68: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Cycle Consistency Loss

G(x) F(G x )x

F G x − x1

Large cycle lossSmall cycle loss

[Zhu*, Park*, Isola, and Efros, ICCV 2017]

DY(G x )

Reconstructionerror

slides credit: Isola / Zhu

Page 69: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

G(x) F(G x )x F(y) G(F x )𝑦

Cycle Consistency Loss

F G x − x1

G F y − 𝑦1

[Zhu*, Park*, Isola, and Efros, ICCV 2017]

DY(G x )

Reconstructionerror

Reconstructionerror

DG(F x )

slides credit: Isola / Zhu

Page 70: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Cycle GAN - Overview

https://junyanz.github.io/CycleGAN/ [Zhu et al. 17.]Prof. Leal-Taixé and Prof. Niessner 79

https://junyanz.github.io/CycleGAN/
Page 71: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Monet’s paintings → photos

slides credit: Isola / Zhu

Page 72: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

slides credit: Isola / Zhu

Page 73: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

slides credit: Isola / Zhu

Page 74: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

Next Lectures

• Next Lectures:– Videos– Neural Rendering– 3D Deep Learning

• Keep working on the projects!

Prof. Leal-Taixé and Prof. Niessner 83

Page 75: GAN Architectures and Conditional GANs · transition between the original and edited projection different degree of image manipulation Editing UI Slide credit Zhu / [Zhu et al. 16]

See you next week

Prof. Leal-Taixé and Prof. Niessner 84


Related Documents
GAN-Leaks: A Taxonomy of Membership Inference Attacks ... · GAN-Leaks: A Taxonomy of Membership Inference Attacks against GANs Dingfan Chen1 Ning Yu2,3 Yang Zhang1 Mario Fritz1 1CISPA

GAN-Leaks: A Taxonomy of Membership Inference Attacks...

Category: Documents
Dosimetric evaluation of synthetic CT generated with GANs ... · artificial intelligence, brain MRI, deep learning, GAN structure, proton therapy, synthetic CT----- This is an open

Dosimetric evaluation of synthetic CT generated with GANs...

Category: Documents
GAN Compression: Efficient Architectures for Interactive … · 2020. 6. 29. · GAN Compression: Efficient Architectures for Interactive Conditional GANs Muyang Li1,3 Ji Lin1 Yaoyao

GAN Compression: Efficient Architectures for Interactive...

Category: Documents
MedGAN CGAN b-GAN ff LAPGAN MPM-GAN …2017/12/09  · NIPS 2017 Workshop: Deep Learning: Bridging Theory and Practice Long Beach, 2017-12-09 Bridging Theory and Practice of GANs

MedGAN CGAN b-GAN ff LAPGAN MPM-GAN …2017/12/09  · NIPS...

Category: Documents
GaN Lamb wave resonators - Electrical Engineering and ...minar/pdf/GaN_Lamb_wave_resonators… · Lamb Wave Dispersion in Gallium Nitride Micromechanical Resonators Haoshen Zhu, Azadeh

GaN Lamb wave resonators - Electrical Engineering and...

Category: Documents
Compositional Pattern Producing GAN · Generative Adversarial Networks (GANs) are a class of methods capable of generating photorealis-tic images [7, 16, 4]. They posit a generative

Compositional Pattern Producing GAN · Generative...

Category: Documents
Gans, Haus-/Graugans...die Diepholzer Gans, die Emdener Gans, die Pommersche Gans und die Rheinische Gans. Die Wildgänse unterteilt man in zwei Gruppen: In die Feldgänse mit zehn

Gans, Haus-/Graugans...die Diepholzer Gans, die Emdener...

Category: Documents
AE-OT-GAN: Training GANs from data specific latent distribution · 2020-01-28 · AE-OT-GAN: Training GANs from data specific latent distribution Dongsheng An1, Yang Guo1, Min Zhang2,

AE-OT-GAN: Training GANs from data specific latent...

Category: Documents
Gans Burgenland 2015

Gans Burgenland 2015

Category: Documents
SPIGAN: PRIVILEGED ADVERSARIAL LEARNING FROM SIMULATION · 2017). In particular, CycleGAN (Zhu et al., 2017) leverages cycle consistency using a forward GAN and a backward GAN to

SPIGAN: PRIVILEGED ADVERSARIAL LEARNING FROM SIMULATION ·....

Category: Documents
Gans Burgenland 2014

Gans Burgenland 2014

Category: Documents
Gans, Gans & Associates HA-Director of Operations-pd .pdf · Gans, Gans & Associates 7445 Quail Meadow Road, Plant City, FL 33565 813-986-4441 . Director of Operations Durham, NC

Gans, Gans & Associates HA-Director of Operations-pd...

Category: Documents