Adapting CutLER for other detectors

Hi, fantastic work on CutLER! From my understanding, CutLER is an approach to train a supervised detector, such as a Cascade Mask R-CNN, in an unsupervised manner using ViT, MaskCut, and DropLoss techniques. This implies that, theoretically, we should be able to replace the R-CNN detector with a single-stage detector like YOLO, provided we can modify the loss function accordingly.
I am planning to implement a few-shot learning approach by training a self-supervised model using the CutLER methodology on YOLOv8/YOLOv10, followed by fine-tuning on a specific dataset. Could you provide any suggestions on how to modify the YOLOv10 loss function to be compatible with the CutLER approach? Thank you!