I have tested your yolov26 models on a drone video I had (which was really low-quality) and they are giving good results. Was especially impressed by your yolov26-nano model which gave better detections than some of the other larger open-source models I tried, though those were trained on coco.
I would also suggest you to look into finetuning and experimenting with the yolo-master model too. Results from that one were closer to yours and they have an experimental sparse sahi implementation which may be useful.
Also, I have noticed that most of your models on huggingface are labeled with Apache 2.0 license. Maybe you should revisit that as both the VisDrone dataset and the Ultralytics models which you have trained on have different and more restrictive licenses.