Conditional GAN

Created December 15, 2020 ยท Updated March 4, 2026

GANs do not have an encoder, so it is not clear how you can condition the generation on another variable.

Authors propose conditioning on labels. Appending one-hot-encoded label vector to noise vector

$$ \min _{G} \max _{D} \mathbb{E}_{x \sim p_{\text {data}}}[\log D(x \mid y)]+\mathbb{E}_{\mathbf{z} \sim p(\mathbf{z})}[\log (1-D(G(\mathbf{z} \mid \boldsymbol{y})))] $$

Discriminator also recieves the label.

cgan

Image to image translation

Conditioning GAN on other images (like edges) to generate relevant images:

im2im-gan

References

  1. Mirza and Osindero, Conditional Generative Adversarial Nets
  2. Isola, Zhu, Zhou, Efros, Image-to-lmage Translation with Conditional Adversarial Networks