Reproduce “Precipitation Downscaling with Spatiotemporal Video Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Prakhar Srivastava, Ruihan Yang, Gavin Kerrigan, Gideon Dresdner, Jeremy McGibbon, Christopher Bretherton, Stephan Mandt
Researchers from the University of California, Irvine and the Allen Institute for AI introduce SpatioTemporal Video Diffusion (STVD), a deep learning framework for precipitation downscaling. STVD generates high-resolution precipitation predictions from coarse-grid inputs, accurately reproducing extreme weather event statistics and fine spatial patterns, addressing limitations of prior deterministic and GAN-based methods.
Key insights
- Conditional video diffusion models provide a robust solution for statistical downscaling, overcoming the mode-averaging problem prevalent in deterministic approaches and the limitations of GANs.
- Incorporating spatio-temporal attention is crucial for leveraging the temporal continuity of precipitation patterns, leading to more physically consistent and realistic downscaled outputs.
- The approach of modeling high-resolution output as a sum of a deterministic initial prediction and a stochastically generated residual proves effective for capturing both average and fine-scale details.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.