A new measure of rarity developed by the researchers adapts the use of K-Nearest Neighbors (KNNs) to represent genuine (training) and synthetic (output) data sets in an image fusion system. Regarding this new analysis method, the authors state:We assume that common samples will be placed closer together, while unique and rare samples will be placed unevenly in the feature space.The result image above shows the smallest nearest neighbor distances (NNDs) and the largest in the StyleGAN architecture trained on FFHQ. For all datasets, the samples with the smallest NNDs show representative and typical images. On the contrary, the samples with the largest NNDs have strong individuality and are significantly different from typical images with the smallest NNDs. Theoretically, by using this new metric as a discriminator, or at least incorporating it into a more sophisticated discriminator architecture, it is possible to guide the generative system from pure imitation to a more inventive algorithm, while preserving the essential coherence of concepts that may be critical for generating authentic images (e.g., "man", "woman", "car", "church", etc.).
English 
Русский