Recent studies have shown remarkable success in image-to-image translation for two domains. However, existing approaches have limited scalability and robustness in handling more than two domains, since different models should be built independently for every pair of image domains. We propose StarGAN, a novel and scalable approach that can perform image-to-image translations for multiple domains using only a single model. Such a unified model architecture allows simultaneous training of multiple datasets with different domains within a single network. Experiments demonstrate that StarGAN generates visually higher quality results compared to existing methods.

Existing methods for image-to-image translation, such as pix2pix and CycleGAN, learn mappings between two specific domains. To handle translations among k domains, these approaches require training k(k-1) separate generators, which is both computationally expensive and unable to leverage training data from other domains. Our key innovation is that the generator takes in as inputs both the image and the target domain information, and learns to flexibly translate the image into the corresponding domain. We additionally introduce a mask vector technique that enables joint training across datasets with different label sets.

Generative Adversarial Networks (GANs) have been applied to various tasks including image generation, super-resolution, and domain adaptation. Conditional GANs condition the generation on additional information like class labels. Image-to-image translation methods such as pix2pix require paired training data, while CycleGAN and DiscoGAN learn from unpaired data using cycle consistency loss. However, these methods are inherently limited to two-domain translation. IcGAN and DIAT attempt multi-domain translation but require separate models or produce lower quality results.

StarGAN uses a single generator G and discriminator D. The generator takes as input an image x and a target domain label c, and produces a translated image G(x, c). The discriminator has two branches: one for real/fake classification (adversarial) and one for domain classification. This design means that a single G learns all possible domain translations, using the target domain label to control which translation to perform. The domain label is concatenated to the input image as additional channels, spatially replicated to match the image dimensions.

The training objective combines four components. (1) Adversarial loss: uses Wasserstein GAN with gradient penalty (WGAN-GP) to produce realistic images. (2) Domain classification loss: ensures the translated image is correctly classified into the target domain, applied to both real images (training D) and fake images (training G). (3) Reconstruction loss: applies cycle consistency — translating from source to target and back should recover the original image, preserving content while only changing domain-specific attributes. The hyperparameters are set to lambda_cls = 1 and lambda_rec = 10.

A key challenge in multi-dataset training is that different datasets have different label sets — for example, CelebA has facial attributes (hair color, gender, age) while RaFD has facial expressions (happy, angry, fearful). StarGAN addresses this with a mask vector approach: the domain label is extended to include all possible labels from all datasets, with a one-hot mask indicating which dataset the label belongs to. Unknown labels are simply set to zero. This enables the first successful multi-domain translation across different datasets in a single model.

Experiments are conducted on CelebA (202,599 facial images, 40 binary attributes) and RaFD (4,824 images, 8 expressions). In AMT user studies for single-attribute transfer on CelebA, StarGAN shows significant preference over baselines: hair color 66.2% vs. CycleGAN's 20.0%, gender 39.1% vs. DIAT's 31.4%, and aging 70.6% vs. IcGAN's 9.2%. On RaFD, the expression classification error is 2.12% for StarGAN versus 5.99% for CycleGAN. Critically, StarGAN requires only 53.2M parameters for all translations, compared to 52.6M x 7 for DIAT and 52.6M x 14 for CycleGAN.

We have proposed StarGAN, a scalable image-to-image translation model capable of multi-domain translations using a single generator-discriminator pair. By incorporating domain labels as input and using mask vectors for multi-dataset training, StarGAN achieves the first successful multi-domain translation across different datasets. The results demonstrate superior image quality through implicit data augmentation from multi-task learning, along with significantly fewer parameters than multi-model approaches. StarGAN represents a step toward more flexible and efficient generative models.