StarGAN v2 is a novel framework designed to address the challenges of image-to-image translation across multiple domains, focusing on both diversity and scalability. The authors propose a single generator that can synthesize diverse images in multiple domains, improving upon existing methods that either lack diversity or require multiple models for each domain. StarGAN v2 introduces a mapping network and a style encoder to learn domain-specific style codes, enabling the generator to produce diverse images reflecting the styles of reference images. The framework is evaluated on the CelebA-HQ dataset and a new animal faces dataset (AFHQ), demonstrating superior visual quality, diversity, and scalability compared to baseline methods. The authors also release the AFHQ dataset, which contains high-quality animal faces with significant inter- and intra-domain variations, to better assess the performance of image-to-image translation models.StarGAN v2 is a novel framework designed to address the challenges of image-to-image translation across multiple domains, focusing on both diversity and scalability. The authors propose a single generator that can synthesize diverse images in multiple domains, improving upon existing methods that either lack diversity or require multiple models for each domain. StarGAN v2 introduces a mapping network and a style encoder to learn domain-specific style codes, enabling the generator to produce diverse images reflecting the styles of reference images. The framework is evaluated on the CelebA-HQ dataset and a new animal faces dataset (AFHQ), demonstrating superior visual quality, diversity, and scalability compared to baseline methods. The authors also release the AFHQ dataset, which contains high-quality animal faces with significant inter- and intra-domain variations, to better assess the performance of image-to-image translation models.