Abstract
Super-resolution (SR) for remote sensing imagery often fails underout-of-distribution (OOD) conditions, such as rare geomorphic features capturedby diverse sensors, producing visually plausible but physically inaccurateresults. We present RareFlow, a physics-aware SR framework designed for OODrobustness. RareFlow's core is a dual-conditioning architecture. A GatedControlNet preserves fine-grained geometric fidelity from the low-resolutioninput, while textual prompts provide semantic guidance for synthesizing complexfeatures. To ensure physically sound outputs, we introduce a multifaceted lossfunction that enforces both spectral and radiometric consistency with sensorproperties. Furthermore, the framework quantifies its own predictiveuncertainty by employing a stochastic forward pass approach; the resultingoutput variance directly identifies unfamiliar inputs, mitigating featurehallucination. We validate RareFlow on a new, curated benchmark of multi-sensorsatellite imagery. In blind evaluations, geophysical experts rated our model'soutputs as approaching the fidelity of ground truth imagery, significantlyoutperforming state-of-the-art baselines. This qualitative superiority iscorroborated by quantitative gains in perceptual metrics, including a nearly40\% reduction in FID. RareFlow provides a robust framework for high-fidelitysynthesis in data-scarce scientific domains and offers a new paradigm forcontrolled generation under severe domain shift.