<p>Image-to-image (I2I) translation transforms an image from a source domain to a target domain while preserving the source content. Many computer vision applications fall within the field of I2I translation, including style transfer, image segmentation, and photo enhancement. The degree to which source image content is preserved during translation varies according to the problem and intended application. From this perspective, we categorize I2I translation tasks into three classes: fully content-preserving, partially content-preserving, and non-content-preserving. For each category, we present representative tasks, datasets, methods, and experimental results. We propose a taxonomy of I2I methods based on model architectures and analyze each category separately. Additionally, we introduce established evaluation criteria in the I2I translation field. Specifically, we analyze nearly 70 different I2I models and introduce more than 10 quantitative evaluation metrics alongside 30 distinct tasks and datasets relevant to I2I translation. Simulation-to-real image translation can be viewed as an application of fully or partially content-preserving unsupervised I2I translation methods. Therefore, we provide a benchmark for sim-to-real translation that can be used to evaluate different approaches. We conclude that due to varying content preservation requirements across applications, careful consideration of this factor is essential when selecting an appropriate I2I model for a specific task.</p>