Research map / Image generation / Super-resolution
Precipitation Downscaling with Spatiotemporal Video Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STVD, a spatiotemporal video diffusion model that combines a deterministic UNet downscaler with a temporally conditioned diffusion model for residuals, to downscale FV3GFS precipitation by 8x.
Problem
Statistical precipitation downscaling needs accurate conditional distributions of high-resolution precipitation given low-resolution inputs (for ensemble averages and extremes), which deterministic and GAN-based super-resolution methods fail to capture, especially across time.
Approach
Decompose downscaling into a deterministic spatio-temporal factorized-attention UNet prediction plus a conditional DDPM that generates residuals for the whole frame sequence jointly, with the two modules trained end-to-end and sharing features.
Key results
- STVD achieves CRPS 1.85e-5, MSE 0.59e-8, EMD 2.49e-6, PE 1.2e-3, and SAE 4.00e-6, best on all metrics versus the baselines (best baseline PSRT: CRPS 2.15e-5, MSE 0.66e-8).
- Removing additional input channels (STVD-single) degrades CRPS to 1.81e-5 is slightly better but MSE, EMD, PE and SAE worsen (EMD 4.64e-6, SAE 6.09e-6), and reducing context to 1 frame (STVD-1) yields EMD 7.19e-6.
- Swin-IR overestimates large precipitation events while the other baselines underestimate extremes; STVD matches the precipitation distribution most closely.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STVD instead of PSRT | FV3GFS precipitation downscaling · FV3GFS validation | CRPS (10^{-5}): 2.15 -> 1.85 | Helped |
| Use STVD instead of PSRT | FV3GFS precipitation downscaling · FV3GFS validation | MSE (10^{-8}): 0.66 -> 0.59 | Helped |
| Use STVD instead of PSRT | FV3GFS precipitation downscaling · FV3GFS validation | EMD (10^{-6}): 4.21 -> 2.49 | Helped |
| Use STVD instead of PSRT | FV3GFS precipitation downscaling · FV3GFS validation | PE (10^{-3}): 3.8 -> 1.2 | Helped |
| Use STVD instead of PSRT | FV3GFS precipitation downscaling · FV3GFS validation | SAE (10^{-6}): 6.24 -> 4.00 | Helped |
| Add the additional climate input channels (ablation) | STVD · FV3GFS validation | MSE (10^{-8}): 0.62 -> 0.59 | Helped |
| Increase context length from 3 frames to 5 frames (ablation) | STVD · FV3GFS validation | CRPS (10^{-5}): 1.96 -> 1.85 | Helped |
| Increase context length from 3 frames to 5 frames (ablation) | STVD · FV3GFS validation | MSE (10^{-8}): 0.68 -> 0.59 | Helped |
| Increase context length from 3 frames to 5 frames (ablation) | STVD · FV3GFS validation | EMD (10^{-6}): 4.94 -> 2.49 | Helped |
| Increase context length from 3 frames to 5 frames (ablation) | STVD · FV3GFS validation | PE (10^{-3}): 2.6 -> 1.2 | Helped |
| Increase context length from 3 frames to 5 frames (ablation) | STVD · FV3GFS validation | SAE (10^{-6}): 4.99 -> 4.00 | No clear effect |
| Increase context length from 1 frame to 5 frames (restore temporal attention) (ablation) | STVD · FV3GFS validation | CRPS (10^{-5}): 2.05 -> 1.85 | Helped |
Surprising findings
- STVD-single has slightly better CRPS (1.81e-5) than the full STVD (1.85e-5) in the reported table, despite being worse on the other metrics.
- Swin-IR (deterministic) performs worst on EMD, PE and SAE, with PE of 23.4e-3, far above other methods.
Limitations
- Requires paired low-resolution and high-resolution data for training.
- Pilot application limited to FV3GFS simulations; distribution shift could degrade performance on new datasets.
- Autoregressive rollout for multiple sequences is left to future work.
Experimental setup
Datasets: FV3GFS
Benchmarks: FV3GFS
Metrics: CRPS MSE Earth Mover Distance 99.999th percentile error spatial autocorrelation error
Baselines: Swin-IR Swin-IR-Diff VRT RVRT PSRT VDM
Builds on: DDPM SR3 v-parameterization DDIM Video Restoration Transformer
Closest papers by meaning
- PrecipDiff: Leveraging Image Diffusion Models to Enhance Satellite-Based Precipitation Observations (AAAI 2025 · similarity 0.89)
- Quantile-Regression-Ensemble: A Deep Learning Algorithm for Downscaling Extreme Precipitation (AAAI 2024 · similarity 0.86)
- Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution (CVPR 2025 · similarity 0.86)
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.86)
- Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet (NeurIPS 2026 · similarity 0.86)
- Adaptive Flow Matching for Resolving Small-Scale Physics (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.