Multi-level inference builds a cascade of models where the output metadata from a parent model (for example, a YOLO detector) drives child models (for example, ResNet classifiers) on cropped regions of interest (ROI). The vvas_xinfer element attaches a tree of inference metadata to each frame: the root represents the frame, each detection is a child node, and each classification result becomes a child of its detection node. In a 2-stage pipeline, YOLO generates bounding boxes, vvas_xinfer crops each box and runs the classifier, and the classifier results are inserted under the corresponding detection in the metadata tree. The attach-ppe-outbuf option stores the pre-processed ROI buffer inside the metadata tree so that subsequent stages can reuse the same crop (improving performance by avoiding repeated crop/resize operations).