HybridFlow: A Hybrid Velocity Generation Framework for Precipitation Nowcasting

Accurate precipitation nowcasting remains a critical challenge in meteorology due to the highly nonlinear and multi-scale dynamics of atmospheric systems. Current generative models often overlook spatiotemporal causal relationships between frames, leading to the loss of local meteorological dynamics, while latent space-based approaches suffer from structural information attenuation that distorts precipitation morphology. To address these limitations, we propose HybridFlow, a hybrid generative framework tailored for precipitation nowcasting. HybridFlow decomposes the generation process into two complementary branches: a global rectified flow branch that ensures large-scale spatiotemporal consistency, and a local sequence flow branch that captures the temporal evolution of fine-scale precipitation processes, thereby balancing global coherence and local accuracy. Furthermore, we develop a Hierarchically-Injected Latent Decoder (HILD), which introduces hierarchical latent-space injection and adaptive feature modulation. Specifically, Cross-Layer Cross-Attention (CLCA) aligns low-resolution meteorological semantics, while Cross-Layer Affine Modulation (CLAM) enhances high-resolution structural fidelity, effectively mitigating information loss during reconstruction. Extensive experiments on the Swedish and MRMS datasets demonstrate that HybridFlow consistently outperforms state-of-the-art baselines across multiple metrics, improving the Critical Success Index (CSI) by 15.2% at 6.3 mm/h and 19.2% at 8 mm/h, while preserving both structural coherence and perceptual realism in the generated precipitation fields.