Animal recognition can boost, e.g., the search for lost pets and the recognition of individual animals from endangered species. These tasks currently rely on invasive methods like microchips, tattoos, or collars, which are neither reliable nor humane for pets. The lack of large datasets focused on animal face recognition hinders training a deep architecture from scratch to this aim. We investigated the use of pre-trained network models with some transfer learning to animal face recognition, also assessing the influence of possible different image quality. The experiments tested two models, FaceNet trained on large databases of human faces, and ViT trained on generic object categories, hypothesizing that the former should have achieved better performance, and further compared these with State of The Art (SOTA). The benchmarks were a dataset of dog images of good quality, and another with groups of endangered primates (lemurs, golden monkeys, and chimpanzees) photographed in more ‘in the wild’ conditions. We compared the achieved results for each dataset with ad-hoc trained deep networks representing SOTA for the two specific recognition problems. As expected, the best performance was achieved with dog faces due to better image quality. Less expected was that ViT, pre-trained on ImageNet object database, overcame both FaceNet and, most noticeably, the SOTA reaching a mean verification accuracy of 92.7% and a rank-1 identification rate of 69.6%, compared with 88.8% and 39.74% respectively for SOTA that uses an ad-hoc architecture. The results with endangered primates are encouraging too, but the performance varies with the animal class and the task (verification or identification) and does not always outperform SOTA.