Real-time defect detection and process monitoring have become essential for ensuring quality in Large-Format Additive Manufacturing (LFAM). If an issue like warping, delamination, or distortion begins during printing, detecting it immediately and in-situ protects against material waste and prevents future consequences such as compromised part integrity, loss of material and time, and significant expenses. In this study, we present a systematic comparison of three advanced computer vision approaches Faster R-CNN, DETR and YOLOV8 for defect identification and classification within LFAM processes. All models are assessed under identical experimental settings using a custom-labeled LFAM defect dataset. The evaluation includes primary indicators such as frames per second (FPS), memory consumption, and feasibility for real-time deployment. In addition, complementary metrics including accuracy, recall, F1-score, confusion matrix evaluation are used to provide deeper insights into classification performance. Our findings reveal distinct trade-offs between processing speed and detection accuracy across the models. F-RCNN excelled in classifying defect types, whereas DETR was better at learning and localizing defects but showed limited classification accuracy. F-RCNN achieved the highest accuracy (F1 = 0.852, AUC = 0.914), DETR failed (F1 = 0), and YOLOv8 performed poorly (F1 = 0.192) due to class confusion. However, YOLOv8 outperformed in efficiency, with minimal memory usage (147 MB), fastest training (0.086 h), and highest inference speed (25.82 FPS). These results indicate that F-RCNN is preferable for precision, while YOLOv8 is better when speed and low resource usage are prioritized.