A new preprint demonstrates that flow matching, a generative AI framework, outperforms leading diffusion models in downscaling coarse 8 km precipitation data to 2 km convective scales over Singapore, delivering superior spatial structure and fractions skill scores across thresholds. The study trains the model on daily precipitation fields from a limited convective-scale domain and benchmarks it directly against CPMGEM, revealing tighter amplitude and structure components in the SAL metric. However, the approach introduces a notable dry bias by underestimating the upper tail of rainfall distributions, a critical flaw for flood risk assessment in tropical urban settings. This work builds on earlier diffusion-based downscaling efforts, such as those exploring score-based models for regional climate, while addressing computational inefficiencies in traditional dynamical methods that struggle with sub-kilometer resolution. Related research on flow matching in image and video domains highlights its potential for capturing fine-grained patterns missed by physics-only simulations, yet the Singapore-focused sample lacks broader validation across diverse climates or multi-year ensembles. Limitations include the preprint status, absence of explicit sample size details beyond the domain, and no peer review, raising questions about generalizability. The technique sharpens expectations for neighborhood-level rain within yearly horizons but requires bias correction to avoid underpreparing for extremes amid intensifying monsoon variability.