[Paper Review] ICR-GAN: Improved Consistency Regularization for GANs 간단한 논문 리뷰


⭐ 본 논문은 CR-GAN의 후속 논문이다. CR-GAN이 real image에 대해서만 augmentation을 했다면, Improved Consistency Regularization 은 real image뿐만 아니라 generated images와 latent space vector, Generator에 대해서도 Augmentation을 한다.

Improved Consistency Regularization

  • bCR-GAN(Balanced Consistency Regularization) : Discriminator - real image와 generated image 둘다 consistency regularization
  • zCR-GAN(Latent Consistency Regularization) : latent vector를 augment하여 이미지를 생성한 후, G(z)G(T(z)) pair에 대해 consistency regularization
  • ICR-GAN(Improved Consistency Regularization) : bCR + zCR 둘다 !

1. bCR-GAN

네이버 AI랩 최윤제님 발표자료

To address the lack of regularization on the generated samples, bCR-GAN introduces balanced consistency regularization (bCR), where a consistency term on the discriminator is applied to both real and generated samples.

  • $L_{\text {fake }}$ term이 추가되었다.
  • 순서대로 adversarial Loss, CR-Loss, bCR-Loss !
\[L_{D} \leftarrow D(G(z))-D(x) \\ L_{\text {real }} \leftarrow\|D(x)-D(T(x))\|^{2} \\ L_{\text {fake }} \leftarrow\|D(G(z))-D(T(G(z)))\|^{2}\]

2. zCR-GAN

네이버 AI랩 최윤제님 발표자료

zCR-GAN introduces latent consistency regularization (zCR), which regularizes the sensitivity of both the generator and discriminator changes in the prior. Given augmented latent codes, the generator is encouraged to be sensitive to the augmentations while the discriminator is encouraged to be insensitive to them.

다음 알고리즘에서 pertubation noise $\Delta z$ 에 따라 다양한 이미지를 만들도록 $L_{\mathrm{gen}}$는 maximize하는 방향으로 학습하고,

\[L_{\mathrm{gen}} \leftarrow-\|G(z)-G(T(z))\|^{2}\]

Discriminator가 Consistency Regularization을 잘 학습하도록 $L_{\mathrm{dis}}$는 minimize하는 방향으로 학습한다.

\[L_{\mathrm{dis}} \leftarrow\|D(G(z))-D(G(T(z)))\|^{2}\]



ICR-GAN이 baseline보다 성능이 좋다