Introduction. The proliferation of fixed-wing Unmanned Aerial Vehicles (UAVs), such as loitering munitions, presents a significant challenge to airspace security. Detection relies heavily on Thermal Infrared (TIR) imaging for 24/7 passive monitoring. A critical challenge in developing these systems is the "Multiscale Approach" problem: tracking a target as it rapidly transitions from a distant sub-pixel dot to a close-range resolved object. This transition is a critical failure point for modern defense systems, known as the "handover" problem in SHORAD. Existing datasets fail to capture this continuous evolution due to the dangers and costs associated with filming air-to-air collision courses. This data vacuum hinders the development of robust Counter-UAS (C-UAS) algorithms, as traditional synthetic data often lacks the thermodynamic fidelity of real sensors. The purpose of the paper is to introduce Gen-Thermal-UAV, a novel synthetic dataset designed to fill this gap, and to propose a "Seed-Driven" methodology utilizing advanced video diffusion models (Gemini Veo 3). This approach aims to generate high-fidelity synthetic videos that maintain authentic sensor characteristics while simulating diverse flight trajectories, enabling the training of end-to-end trackers robust to extreme scale changes. Methodology. We employed a Seed-Driven Generative AI pipeline. Instead of generating data from scratch, we used Image-to-Video generation anchored by two real thermal images: one far-field (a blurry dot) and one near-field (a resolved plane). This approach ensures thermodynamic fidelity, as the diffusion model propagates the real sensor noise, blur, and heat signatures present in the seed images along realistic flight paths. The diffusion model predicts the motion of the pixel distribution, effectively "hallucinating" the preservation of physics rather than relying on low-fidelity rasterization. A structured prompt engineering taxonomy was developed to constrain the generative model to scientific consistency. The resulting videos were automatically annotated using the Segment Anything Model 2 (SAM 2), leveraging temporal consistency for zero-shot labeling, validated by an 85% confidence filter. Results. Gen-Thermal-UAV comprises 220 videos (1,760 seconds, approx. 42,000 frames) at 720p resolution, depicting air-to-air fixed-wing engagement scenarios. It is the first dataset to capture the continuous dot-to-object transition in the thermal domain. A comparative analysis confirms its unique position at the intersection of Thermal modality, Air-to-Air platform, and Extreme multiscale dynamics, distinguishing it from benchmarks like AOT, HIT-UAV, and Anti-UAV410. Conclusions. Gen-Thermal-UAV addresses a high-value gap in the computer vision landscape for C-UAS applications. The verified methodology demonstrates that Generative AI, when constrained by real-world seed data, can produce physics-compliant training data for dangerous or rare scenarios. This work not only provides a crucial benchmark but also establishes a reproducible protocol for generating and auto-labeling synthetic data, democratizing access to "edge case" training and facilitating rapid adaptation to emerging threats. Keywords: thermal infrared tracking, unmanned aerial vehicles, synthetic data generation, video diffusion models, multiscale detection, computer vision.
Smolin et al. (Mon,) studied this question.