Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (CycleGAN) from UC Berkeley (pix2pix upgrade)

&

Learning to Discover Cross-Domain Relations with Generative Adversarial Networks (DiscoGAN) from SK T-Brain

DiscoGAN & CycleGAN

• Almost Identical concept.
• DiscoGAN came 15 days earlier. Low resolution ($64 \times 64$)
• CycleGAN has better qualitative results ($256 \times 256$) and quantative experiments.

Difference from DTN

• No $f$-constancy. Do not need pre-trained context encoder
• Only need dataset $S$ and $T$ by proposing cycle-consistency

DiscoGAN

without cross domain matching, GAN has mode collapse

learn projection to mode in domain $B$, while two domains have one-to-one relation

Typical GAN issue: Mode collapse

• top is ideal case, bottom is mode collapse failure case
• Toy problem of 2-dim Gaussian mixture model
• 5 modes of domain A to 10 modes of domain B

• GAN, GAN + const show injective mapping & mode collapse
• DiscoGAN shows bijective mapping & generate all 10 modes of B.

proposed DiscoGAN

CycleGAN has similar contribution on this point

Results

codes and more results in

• https://github.com/SKTBrain/DiscoGAN
• https://github.com/carpedm20/DiscoGAN-pytorch

CycleGAN

Use more GAN techniques: LSGAN, use image buffer of previous generated samples

failure case

CycleGAN demonstrates more experiments!
project page : https://junyanz.github.io/CycleGAN/
code available with Torch and PyTorch