KAIST Proposes PROLA: Low-Rank Adaptation for Spatiotemporal Weather Downscaling

MLVLab · hf · 2026-11-19

KAIST researchers introduce PROLA (Principal-Orthogonal Low-rank Adaptation) for Predictive Spatiotemporal Weather Downscaling (PSWD)—jointly refining spatial and temporal resolution while forecasting high-res trajectories. The framework splits prediction into spatial and temporal stages over a shared pretrained backbone with fixed numerical scaffolds and lightweight residual heads. PROLA allocates a fixed low-rank budget between principal directions and their orthogonal complement, and rescales rank-1 optimizer updates using gradient signals. It outperforms representative LoRA baselines on PSWD and generalizes across diverse datasets and tasks.

Original post →

More from Infra

Infra channel →