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 enforce F(G(X)) ≈ X (and vice versa). Results on several tasks demonstrate the effectiveness of our approach, including collection style transfer, object transfiguration, season transfer, and photo enhancement.

Consider the task of translating a photograph into the style of a Monet painting. Obtaining paired training data — the same scene photographed and painted by Monet — is impractical. Yet we can gather a collection of Monet paintings and a collection of landscape photographs. We present a method that can "learn to do the same: capturing special characteristics of one image collection and figuring out how these characteristics could be translated into the other". The key challenge is that adversarial losses alone are under-constrained: a network can map any input to any valid output in the target distribution, causing mode collapse. Our solution is to introduce cycle consistency: if we translate from X to Y and back to X, we should arrive where we started. This structural assumption significantly reduces the space of possible mappings.

Our work builds on pix2pix (Isola et al.) for image-to-image translation but removes the paired data requirement. Prior unpaired methods include CoGAN (using weight-sharing), BiGAN/ALI (using shared embeddings), and SimGAN (using self-regularization). The idea of cycle consistency — using transitivity as regularization — has a long history in visual tracking, 3D structure estimation, and machine translation (back-translation). Unlike neural style transfer (Gatys et al.), which matches Gram matrices between single image pairs, CycleGAN learns collection-level mappings that capture the overall style of an entire domain rather than mimicking a single exemplar.

We define two mapping functions G: X → Y and F: Y → X, with associated discriminators D_Y and D_X. For the mapping G, the adversarial loss is: L_GAN(G, D_Y, X, Y) = E_y[log D_Y(y)] + E_x[log(1 - D_Y(G(x)))]. G tries to generate images G(x) that look similar to images from domain Y, while D_Y aims to distinguish between translated samples G(x) and real samples y. An analogous adversarial loss is applied to F: Y → X with discriminator D_X. However, with large enough capacity, a network can map the same set of input images to any random permutation of images in the target domain, any of which could match the target distribution. Thus, adversarial losses alone cannot guarantee that the learned function can map an individual input x_i to a desired output y_i.

To further reduce the space of possible mapping functions, we argue that the learned mappings should be cycle-consistent. For each image x from domain X, the image translation cycle should bring x back to the original: x → G(x) → F(G(x)) ≈ x (forward cycle consistency). Similarly, for each image y from domain Y: y → F(y) → G(F(y)) ≈ y (backward cycle consistency). The cycle consistency loss is defined as: L_cyc(G, F) = E_x[||F(G(x)) - x||_1] + E_y[||G(F(y)) - y||_1], using the L1 norm. The full objective combines adversarial and cycle losses: L(G, F, D_X, D_Y) = L_GAN(G, D_Y, X, Y) + L_GAN(F, D_X, Y, X) + lambda * L_cyc(G, F), where lambda = 10. Training can be viewed as training two "autoencoders" F(G): X → X and G(F): Y → Y, with the constraint that the intermediate representation passes through the other domain.

The generator architecture is based on Johnson et al.'s style transfer network, with 3 convolutions, 9 residual blocks (for 256x256+), and 2 transposed convolutions, using instance normalization. The discriminator uses a 70x70 PatchGAN that classifies overlapping 70x70 patches as real or fake, rather than the full image. For training stability, we use least-squares loss instead of the standard negative log-likelihood GAN loss, and maintain an image buffer of 50 previously generated images to reduce oscillation. Training uses Adam optimizer with learning rate 0.0002, batch size 1, 100 epochs at constant rate followed by 100 epochs with linear decay.

We evaluate across multiple tasks. In AMT perceptual studies (25 participants per algorithm), CycleGAN achieves 26.8% "real vs. fake" confusion rate for maps-to-photos (compared to CoGAN's 0.6%, SimGAN's 0.7%, and pix2pix's 33.9% with paired data). On Cityscapes labels-to-photos, CycleGAN achieves per-pixel accuracy 0.52 and class IOU 0.11, outperforming all unpaired baselines though still behind pix2pix (0.71, 0.18). Ablation studies confirm that "both terms [adversarial and cycle] are critical to our results": GAN loss alone produces mode collapse, and cycle loss alone produces unrealistic results. Applications include Monet/Van Gogh/Cezanne/Ukiyo-e style transfer, horse-to-zebra, summer-to-winter Yosemite, and smartphone-to-DSLR photo enhancement. An additional identity mapping loss L_identity was introduced for tasks requiring color preservation.

The method "often succeeds" on color and texture changes but has "little success" on geometric transformations. For instance, dog-to-cat translation makes minimal changes because geometric modifications are inherently difficult for the generator architecture. Horse-to-zebra with riders fails on unseen configurations (training data did not include horseback riding). The generator architecture is "tailored for good performance on appearance changes" but not structural ones. There remains a persistent gap between paired (pix2pix) and unpaired (CycleGAN) methods. The authors note that "handling more varied and extreme transformations, especially geometric changes, is an important problem for future work", while emphasizing that "in many cases completely unpaired data is plentifully available and should be made use of".