From YOLOv1 to YOLO26: A Unified Comparative Review of the Evolution, Architecture, and Applications of YOLO-Based Object Detection Frameworks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Object detection has evolved from classical feature-based approaches to deep learning-based frameworks, with the YOLO (You Only Look Once) family emerging as one of the most widely adopted paradigms for real-time object detection. Despite rapid architectural evolution from YOLOv1 to YOLO26, the literature remains fragmented and lacks a unified synthesis integrating architectural innovations, quantitative benchmarking, and domain-specific deployment considerations. This review presents a comprehensive longitudinal analysis of the YOLO lineage, spanning the evolution from Darknet-based backbones to hypergraph-enhanced correlation modeling in YOLOv13 and end-to-end NMS-free architectures in YOLO26, as well as the transition from anchor-based prediction to anchor-free, decoupled, and end-to-end detection heads. We synthesize source-reported metrics, including AP, FPS, FLOPs, and parameter counts, across both canonical and domain-specific benchmarks, while explicitly treating the resulting comparison as a cross-source synthesis rather than a hardware-normalized re-benchmark. Particular attention is given to developments reported during 2025–2026. Under their respective source protocols, YOLOv13 reports a 3.0-AP-point improvement over YOLO11-N (38.6 to 41.6 AP), whereas YOLO26 reports up to 43% faster CPU inference than YOLO11. Because both releases are relatively recent, these findings are presented as source-reported results and are accompanied by an explicit maturity caveat informed by emerging independent evaluations. Across seven application domains, we examine how architectural design choices relate to operational constraints, identify persistent challenges in domain generalization, small-object localization, and open-set detection, and propose future research directions centered on contextual representation, deployment-aware training, data-efficient learning, and reproducible, sustainability-aware benchmarking.