YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family
YOLO-PEFT turns adapter placement for YOLO-style detectors into a rule-checked planning step before training.
The paper says generic PEFT methods from language models can fail on real-time detectors because YOLO-family graphs have operator and deployment constraints Transformers do not. Its planner labels detector modules, checks validity and budget predicates, and either emits target modules or refuses before training. In the reported VOC setup, planner-selected RS-LoRA beats Full-SFT mAP50-95 on YOLO11s and YOLO12s, while RT-DETR-L results support refusing PEFT in favor of Full-SFT. A YOLO11 audit found LoRA cut peak training memory by 43.9%, with training taking 1.72x longer. HF Daily Papers' note
The paper says generic PEFT methods from language models can fail on real-time detectors because YOLO-family graphs have operator and deployment constraints Transformers do not. Its planner labels detector modules, checks validity and budget predicates, and either emits target modules or refuses before training. In the reported VOC setup, planner-selected RS-LoRA beats Full-SFT mAP50-95 on YOLO11s and YOLO12s, while RT-DETR-L results support refusing PEFT in favor of Full-SFT. A YOLO11 audit found LoRA cut peak training memory by 43.9%, with training taking 1.72x longer. HF Daily Papers' note
score 4