Conventional aesthetic frameworks, largely derived from human-dominated environments, can undervalue the Earth's remaining wildernesses and often portray them as visually barren. Here, we operationalise wilderness aesthetics as a spatially explicit cultural ecosystem service indicator and examine a paradox on the Tibetan Plateau: ecological adversity is strongly associated with aesthetic power. We compile a plateau-specific image dataset (Landscape-SOB, n = 19,522) and train an attention-enhanced ResNet-50 + SE model (overall accuracy = 87.0%) to map eight aesthetic types, including "Wilderness", from social-media imagery. Treating model-derived aesthetic intensities as county-level spatial indicators, we combine spatial autocorrelation analysis with GeoDetector and GWR to quantify associations, interaction effects and spatial heterogeneity between aesthetics and landscape attributes. Results reveal a pronounced southeast-northwest dichotomy: forested southeast is dominated by "Graceful" and "Secluded" aesthetics, whereas the arid northwest is characterised by "Wilderness" and "Expansive" types, where harsh environmental conditions outweigh vegetation greenness, challenging the "green-is-beautiful" heuristic. In this "aesthetics of adversity", low vegetation cover (NDVI <0.2) and high relief variability (RV > 0.4) form a core wilderness signal, governed by non-linear synergies and marked spatial non-stationarity. Our work advances landscape appraisal from a subjective pursuit toward a predictive, spatially explicit science, providing a quantitative basis for conserving the aesthetic capital of the planet's last great wilderness landscapes.