Diagnosing Aerial-View Object Detectors with Foundational Image Generative Models

Recent advances in large-scale image generative models en-able photorealistic scene synthesis with controllable attributes. Beyonddata augmentation, their potential as diagnostic tools for trained visionsystems remains unexplored in the aerial and remote sensing domains.We introduce a synthetic diagnostic framework for aerial-view vehicle de-tection that combines text-guided generation, attribute-controlled edit-ing, and automated attribute verification to construct a controllable syn-thetic testbed. This enables fine-grained evaluation of pretrained detec-tors under diverse scene types and environmental conditions that aredifficult to isolate in real datasets. Across three detection architecturesand three real aerial datasets, synthetic scene-wise performance trendsclosely match real-world weaknesses. Guided by these diagnostics, tar-geted supplementation with small real datasets from the identified weakcategories yields improvements of up to 13% AP50 while requiring sub-stantially fewer additional samples than non-targeted augmentation. Ourresults show that controlled synthetic probing can predict real-domainperformance gaps and guide efficient data collection. The proposed diag-nostic framework is modular and can incorporate alternative generativeor vision-language models as capabilities evolve. Our code and datasetsare available here: humansensinglab.github.io/AVODDiag/