Generation Models

CycleGAN (ICCV’2017)


Image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between an input image and an output image using a training set of aligned image pairs. However, for many tasks, paired training data will not be available. We present an approach for learning to translate an image from a source domain X to a target domain Y in the absence of paired examples. Our goal is to learn a mapping G:X→Y such that the distribution of images from G(X) is indistinguishable from the distribution Y using an adversarial loss. Because this mapping is highly under-constrained, we couple it with an inverse mapping F:Y→X and introduce a cycle consistency loss to push F(G(X))≈X (and vice versa). Qualitative results are presented on several tasks where paired training data does not exist, including collection style transfer, object transfiguration, season transfer, photo enhancement, etc. Quantitative comparisons against several prior methods demonstrate the superiority of our approach.

Results and models

We use FID and IS metrics to evaluate the generation performance of CycleGAN.

Method FID IS Download
official facades 123.626 1.638 -
ours facades 118.297 1.584 model | log
official facades-id0 119.726 1.697 -
ours facades-id0 126.316 1.957 model | log
official summer2winter 77.342 2.762 -
ours summer2winter 76.959 2.768 model | log
official winter2summer 72.631 3.293 -
ours winter2summer 72.803 3.069 model | log
official summer2winter-id0 76.773 2.750 -
ours summer2winter-id0 76.018 2.735 model | log
official winter2summer-id0 74.239 3.110 -
ours winter2summer-id0 73.498 3.130 model | log
official horse2zebra 62.111 1.375 -
ours horse2zebra 63.810 1.430 model | log
official horse2zebra-id0 77.202 1.584 -
ours horse2zebra-id0 71.675 1.542 model | log
official horse2zebra 138.646 3.186 -
ours zebra2horse 139.279 3.093 model | log
official horse2zebra-id0 137.050 3.047 -
ours zebra2horse-id0 132.369 2.958 model | log
official average 95.935 2.444 -
ours average 95.102 2.427 -

Note: With a larger identity loss, the image-to-image translation becomes more conservative, which makes less changes. The original authors did not say what is the best weight for identity loss. Thus, in addition to the default setting, we also set the weight of identity loss to 0 (denoting id0 ) to make a more comprehensive comparison.


  title={Unpaired image-to-image translation using cycle-consistent adversarial networks},
  author={Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A},
  booktitle={Proceedings of the IEEE international conference on computer vision},

Pix2Pix (CVPR’2017)


We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. This makes it possible to apply the same generic approach to problems that traditionally would require very different loss formulations. We demonstrate that this approach is effective at synthesizing photos from label maps, reconstructing objects from edge maps, and colorizing images, among other tasks. As a community, we no longer hand-engineer our mapping functions, and this work suggests we can achieve reasonable results without hand-engineering our loss functions either.

Results and models

We use FID and IS metrics to evaluate the generation performance of pix2pix.

Method FID IS Download
official facades 119.135 1.650 -
ours facades 127.792 1.745 model | log
official maps-a2b 149.731 2.529 -
ours maps-a2b 118.552 2.689 model | log
official maps-b2a 102.072 3.552 -
ours maps-b2a 92.798 3.473 model | log
official edges2shoes 75.774 2.766 -
ours edges2shoes 85.413 2.747 model | log
official average 111.678 2.624 -
ours average 106.139 2.664 -

Note: we strictly follow the paper setting in Section 3.3:

At inference time, we run the generator net in exactly the same manner as during the training phase. This differs from the usual protocol in that we apply dropout at test time, and we apply batch normalization using the statistics of the test batch, rather than aggregated statistics of the training batch.

i.e., use model.train() mode, thus may lead to slightly different inference results every time.


  title={Image-to-image translation with conditional adversarial networks},
  author={Isola, Phillip and Zhu, Jun-Yan and Zhou, Tinghui and Efros, Alexei A},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
Read the Docs v: stable
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.