It's possible to adapt ResNets for unsupervised learning. A popular approach is to use self-supervised learning techniques, where the model is initially trained to predict some aspect of the data (e.g., predicting a missing part of an image) without explicit labels. This is similar to masked language modeling.