Which technique involves training models on unpaired datasets for image conversion?

The technique that involves training models on unpaired datasets for image conversion is CycleGAN. This model is particularly designed to learn a mapping between two different domains without requiring paired examples. For instance, if you wanted to convert images of horses to zebras, CycleGAN can learn to perform this transformation effectively even when the training dataset contains images of horses and zebras that are not correspondingly paired.

CycleGAN achieves this by employing a cycle consistency loss, which ensures that if an image is transformed from domain A to domain B and then back to domain A, the result should be the original image. This approach allows the model to learn not just direct mappings between the two domains, but also how to preserve the inherent properties of the images during the transformation. This capability is crucial for applications like style transfer, where maintaining some characteristics of the original images is essential.