Two-way GAN

Cycle-consitenty Loss is good for color and texture, but not very succesfull on shape change. For transfer with shape, could check UNIT and its variants

CycleGAN

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (ICCV 2017) by Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros
Project| Torch | pyTorch | CVPR2018 slides
Could be applyed on any unapired datasets (better if two datasets share similar visual content)
Adversarial Loss

\[\begin{split}\text{for }G: X \to Y{ and } D_Y, \mathfrak{L}_{GAN}(G, D_Y, X, Y) & =\mathbb{E}_{y \sim p_{data}(y)}[log D_Y(y)] \\ & +\mathbb{E}_{x \sim p_{data}(x)}[log (1-D_Y(G(x))] \\ min_G max_{D_Y} \mathfrak{L}_{GAN}(G, D_Y, X, Y)\\\end{split}\]

Cycle Consistency Loss

\[\begin{split}\mathfrak{L}_{cyc}(G,F) & = \mathbb{E}_{x \sim p_{data}(x)} [||F(G(x))-x||_1]] \\ & + \mathbb{E}_{y \sim p_{data}(y)} [||F(G(y))-y||_1]] \\\end{split}\]

Full Objective

\[\begin{split}\mathfrak{L}(G, F, D_x, D_Y) & = \mathfrak{L}_{GAN}(G, D_Y, X, Y) \\ & + \mathfrak{L}_{GAN}(F, D_X, X, Y) \\ & + \mathfrak{L}_{cyc}(G,F) \\ G^*, F^* = arg min_{G, F} max_{D_X, D_Y} \mathfrak{L}(G, F, D_x, D_Y)\end{split}\]

my testing

testing cycleGAN good at color and texture changes, bad at shape

Other two-way GAN

Dual GAN: DualGAN: Unsupervised Dual Learning for Image-to-Image Translation (ICCV 2017)
DiscoGAN: Learning to Discover Cross-Domain Relations with Generative Adversarial Networks (ICML 2017)

Augmented CycleGAN

Augmented CycleGAN: Learning Many-to-Many Mappings from Unpaired Data (ICML 2018)
pyTorch (Python2, pyTorch 0.3) | Theano re-implementation
Apart from generator, also have 2 encoders \(E_A: A \times B → Z_a, E_B: B \times A → Z_b\) which enable optimization of cycle-consistency with stochastic, structured mapping

Paired CycleGAN

PairedCycleGAN: Asymmetric Style Transfer for Applying and Removing Makeup (CVPR 2018)
Could apply specified style from input_reference to input_source, as a one-to-many transformation

pre-train makeup removal function F(many-to-one) with CycleGAN first, then alternate the training of makeup transfer function G (one-to-many)
Note: Consider as conditional GAN with (input source + input reference) as conditions

ComboGAN

ComboGAN: Unrestrained Scalability for Image Domain Translation (CVPR 2018)
encoder-decoder pairs that share the latent coding.

StarGAN

StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation (CVPR 2018)
pyTorch code input mask vector, an one-shot label, the target domain as second input condition

StarGANv2

StarGAN v2: Diverse Image Synthesis for Multiple Domains (CVPR 2020)
stargan-v2

Comparison of above 4 conditional variants

GAN	condition apply to	the amount of generator/encoder	input
Augmented CycleGAN	within domain	2 generators	source image + encoding
Paired CycleGAN	within domain	2 generators	source image + reference image
UNIT, ComboGAN	cross-domains	encoder-decoder pairs	source image (to target encoder)
StarGAN	cross-domains	1 generators	source image + label of target domain