Research map / Probabilistic and scientific modeling / Molecules and proteins
DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting
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 DiffCast, a residual diffusion framework that wraps any deterministic spatio-temporal backbone with a diffusion model (GTUNet) that models local stochastic residuals for precipitation nowcasting, trained end-to-end.
Problem
Precipitation nowcasting from radar echo sequences: deterministic models produce blurry forecasts with faded high-value echoes, while probabilistic/diffusion models produce realistic textures but inaccurate rainfall positions.
Approach
Decompose precipitation evolution into a global deterministic motion trend (from any backbone predictor) plus a local stochastic residual modeled autoregressively by a segment-level residual diffusion model conditioned on the backbone's hidden motion features, trained jointly end-to-end with a deterministic loss.
Key results
- On SEVIR, DiffCast with SimVP raises CSI from 0.2662 to 0.3077 (+15.59%) and cuts LPIPS from 0.3369 to 0.1812 (+53.70% LPIPS gain).
- DiffCast improves all four backbones on the four radar datasets, with CSI and HSS gains reported roughly 2% to 20% in the abstract-level summary and LPIPS improvements of 16.9% to 66.7%.
- End-to-end training outperforms two-stage (frozen backbone) training on SEVIR; e.g., DiffCast SimVP reaches CSI 0.3077 vs 0.2739 when frozen.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Equip SimVP with DiffCast residual diffusion | SimVP · SEVIR | CSI: 0.2662 -> 0.3077 | Helped |
| Equip SimVP with DiffCast residual diffusion | SimVP · MeteoNet | CSI: 0.3346 -> 0.3511 | Helped |
| Equip MAU with DiffCast residual diffusion | MAU · Shanghai Radar | CSI: 0.3996 -> 0.4089 | Helped |
| Equip MAU with DiffCast residual diffusion | MAU · CIKM | CSI: 0.2936 -> 0.3158 | Helped |
| Set α=1, removing the deterministic loss (ablation) | DiffCast SimVP · SEVIR | CSI: 0.3077 -> 0.2430 | Hurt |
| Set α=0, removing the denoising loss (ablation) | DiffCast SimVP · SEVIR | CSI: 0.3077 -> 0.2578 | Hurt |
| Freeze the separately trained SimVP backbone instead of end-to-end training (ablation) | DiffCast SimVP · SEVIR | CSI: 0.3077 -> 0.2739 | Hurt |
| Freeze the separately trained Earthformer backbone instead of end-to-end training (ablation) | DiffCast Earthformer · SEVIR | CSI: 0.2823 -> 0.2622 | Hurt |
| Freeze the separately trained PhyDnet backbone instead of end-to-end training (ablation) | DiffCast PhyDnet · SEVIR | CSI: 0.2757 -> 0.2603 | Hurt |
| Remove GlobalNet (ablation) | DiffCast SimVP · SEVIR | CSI: 0.3077 -> 0.2719 | Hurt |
| Remove GlobalNet (ablation) | DiffCast Earthformer · SEVIR | CSI: 0.2823 -> 0.2558 | Hurt |
| Remove GlobalNet (ablation) | DiffCast PhyDnet · SEVIR | CSI: 0.2757 -> 0.2648 | Hurt |
Surprising findings
- Removing the deterministic loss (alpha=1) makes DiffCast SimVP worse than plain SimVP on CSI (0.2430 vs 0.2662 baseline), showing the deterministic component is necessary.
- On MeteoNet, several DiffCast variants (e.g., with MAU, PhyDnet) show slightly lower SSIM than their backbones, so visual-quality gains are not universal.
Limitations
- Experiments downscale all datasets to 128x128 due to computational limits.
- Performance degrades with lead time and the approach relies on a chosen deterministic loss weight alpha=0.5 tuned by ablation.
Experimental setup
Datasets: SEVIR MeteoNet Shanghai Radar CIKM AI Lab precipitation
Benchmarks: SEVIR MeteoNet Shanghai Radar CIKM
Metrics: CSI CSI-pool4 CSI-pool16 HSS LPIPS SSIM
Baselines: SimVP Earthformer MAU ConvGRU PhyDnet MCVD PreDiff STRPM
Builds on: DDPM DDIM MCVD SimVP Earthformer
Closest papers by meaning
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.89)
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.89)
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.88)
- Integrating Multi-Source Data for Long Sequence Precipitation Forecasting (AAAI 2025 · similarity 0.88)
- AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting (CVPR 2025 · similarity 0.87)
- High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation (ICLR 2025 · similarity 0.87)
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.