cs.LGOct 8, 2026
SaveJust Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion
Abstract
Generative diffusion models are well-suited for probabilistic precipitation nowcasting, but existing approaches often rely on separately trained compression or deterministic forecasting components and remain costly at inference due to iterative denoising. We introduce Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model which addresses both issues by forecasting directly in radar space and enabling few-step generation. Radar-space modeling greatly simplifies training and inference and eliminates uncertainty arising from lossy compression. JWS combines Masked Asynchronous Diffusion, a timestep-sampling scheme that preserves clean context while adapting diffusion training to high-dimensional spatio-temporal data, with a simple scoring-rule objective that aligns training with probabilistic forecasting and unlocks few-step generation. On the SEVIR and MeteoNet benchmarks, JWS achieves state-of-the-art probabilistic forecasting performance at reduced training and inference cost. Even our smallest model remains competitive using substantially fewer parameters and more than 17x faster inference.
Figures & tables
Figure 1: Nowcasting made easy. We introduce JWS, an end-to-end, data-space nowcasting model, greatly simplifying training and inference while being more efficient to train than the state-of-the-art FREUD model at comparable CRPS.
Figure 2: Forecast ensembles in pixel space. Left: FREUD ( Schusterbauer et al., 2026b ) samples multiple reconstructions per forecast to represent reconstruction uncertainty. Right: JWS directly samples radar forecasts with a single model, removing the decoding stage.
Figure 3: Training corruption. (a) Framewise is sampled around a shared sequence-level center , retaining variation across frames while keeping context frames clean. (b) Dimensionality-aware timestep shifting increases corruption for larger spatiotemporal inputs. (c) Higher-resolution radar fields retain more structure at fixed corruption ( ), motivating stronger noise.
Figure 4: HDiT backbone.
Figure 5: Distributional Diffusion Models. Distributional Diffusion Models ( De Bortoli et al., 2025 ) learn the posterior distribution, enabling sampling with coarser diffusion time discretization.
Table 4: Training objective. Distributional training ( De Bortoli et al., 2025 ) enables few-step inference, but CRPS-only degrades at high NFE. sDDM outperforms iDDM ( Martorella et al., 2026 ) .
Figure 6: Inference time scaling. Left: JWS scales with ensemble size and NFEs. Right: JWS outperforms Flow Matching, Shortcut models, and FGN in matched comparisons using a single NFE.
| Method | Stages | Param | CRPS | SSIM | HSS | CSI | RI ⋆ | |
| Deterministic | ConvLSTM ( Shi et al., 2015 ) (NeurIPS ’15) | 1 | 14M | 0.0264 | 0.7749 | 0.5232 | 0.4102 | – |
| PredRNN ( Wang et al., 2017 ) (TPAMI ’22) | 1 | 47M | 0.0271 | 0.7497 | 0.5192 | 0.4045 | – | |
| PhyDNet ( Guen & Thome, 2020 ) (CVPR ’20) | 1 | 14M | 0.0253 | 0.7649 | 0.5311 | 0.4198 | – | |
| SimVP ( Gao et al., 2022a ) (CVPR ’22) | 1 | 16M | 0.0259 | 0.7772 | 0.5280 | 0.4153 | – | |
| EarthFormer ( Gao et al., 2022b ) (NeurIPS ’22) | 1 | 9M | 0.0251 | 0.7756 | 0.5411 | 0.4310 | – | |
| Generative | NowcastNet ( Zhang et al., 2023 ) (Nature ’23) | 1 | 35M | 0.0283 | 0.5696 | 0.5365 | 0.4152 | – |
Table 5: SOTA nowcasting on SEVIR. JWS achieves superior CRPS, SSIM, HSS, and RI on SEVIR. model trained at for better results. metric computed with our pipeline.
| Method | Stages | Param | CRPS | SSIM | HSS | CSI | RI ⋆ | |
| D. | EarthFormer ( Gao et al., 2022b ) (NeurIPS ’22) | 1 | 9M | 0.0224 | – | – | 0.2831 | – |
| Generative | NowcastNet ( Zhang et al., 2023 ) (Nature ’23) | 1 | 35M | 0.0277 | – | – | 0.2955 | – |
| PreDiff ( Gao et al., 2023 ) (NeurIPS ’23) | 2 | 105M | 0.0197 | – | – | 0.2546 | – | |
| CasCast ( Gong et al., 2024 ) (ICML ’24) | 3 | 402M | 0.0180 | – | – | 0.3156 | – | |
| FREUD ( Schusterbauer et al., 2026b ) (CVPR ’26) | 2 | 521M | 0.0193 | 0.7312 | 0.2082 | 0.1417 | – | |
| JWS-T/32 ( ours ) | 1 | 9M | 0.0143 | 0.7887 | 0.4026 | 0.2851 | 0.2376 | |
Table 6: MeteoNet nowcasting. JWS achieves state-of-the-art nowcasting on MeteoNet. JWS-T outperforms larger models in CRPS. metric computed with our pipeline.
Figure 9: Deterministic conditioning. Conditioning on a deterministic forecast as explored by CasCast ( Gong et al., 2024 ) improves CSI and HSS, but consistently degrades calibration.
Figure 10: Zero-shot evaluation on SEVIR-Long. JWS shows superior zero-shot extrapolation performance to 180 min into the future ( frames), while also better calibrated.
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Parameter | JWS-T | JWS-S | JWS-B | JWS-L |
| Parameters (M) | 9 | 24 | 94 | 340 |
| High-resolution layers (down & up) | 2 | 2 | 2 | 2 |
| Bottleneck | 6 | 8 | 8 | 20 |
| High resolution hidden dimension | 192 | 256 | 512 | 768 |
| Bottleneck hidden dimension | 256 | 384 | 768 | 1024 |
| Attention heads (HiRes/Bottleneck) | 3/4 | 4/6 | 8/12 | 12/16 |
Table 7: Hyperparameters. Settings of training runs for JWS across model size.
Figure 13: conditioning.
Figure 14: FLOPs vs. Performance. Our JWS-B, JWS-S, and JWS-T models are significantly more compute and parameter efficient while achieving superior nowcasting performance.
Figure 15: Performance over lead time. JWS-B outperforms CasCast and FREUD across the entire lead time range for CRPS and SSIM, while achieving competitive performance for HSS.
Figure 16: Calibration compared to FlowCast. FlowCast ( Ribeiro & Pucer, 2026 ) shows substantially stronger overconfidence than JWS especially for long lead time.
Figure 17: Patch Size Ablation. Varying patch sizes at both hierarchy levels changes the token count and thus increases FLOPs. More tokens/FLOPs improve CRPS, CSI, and HSS.
Figure 18: Model scalability. Our smallest JWS-T model already outperforms FREUD LSM-L ( Schusterbauer et al., 2026b ) . With increased model scale, our model shows superior results.
Figure 19: Overfitting. Training large (L) models directly on full resolution data leads to strong overfitting. Pretraining on downsampled data first mitigates this issue and preserves scalability.
Figure 20: Data scaling. JWS-L overfits on reduced training data, while increased data diversity shows stable training and clearer gains from model scaling.
Figure 21: Confidence-aware priors. Using cleaner deterministic priors for early frames and noisier priors for later frames improves CRPS, calibration, and CSI over uniform prior strength, improving forecast calibration at longer lead times.
Figure 22: Different solvers. Standard Euler sampling outperforms Heun and SDE integration. While DDM sampling achieves superior HSS and SSIM, Euler performs best for CRPS.
Figure 23: Training particles. Increasing training particles does not improve performance.
Figure 24: Expansion layer. Expanding the ensemble earlier does not yield improved performance.
Figure 25: Small-scale verification. We again find no gain using more particles or earlier expansion.
| Num. Samples | CRPS | SSIM | HSS | CSI | RI |
| 64 | |||||
| 128 | |||||
| 256 | |||||
| 512 | |||||
| 1024 |
Table 12: Metric sensitivity. Results report mean standard deviation across eight evaluation seeds. Mean results are close to large evaluation runs even for small sample sizes.
Figure 26: Qualitative forecast. We show three ensemble members over the full 60 min lead time range. Zoom shows a high-precipitation region with high variability.
Figure 27: Qualitative forecast. Three ensemble members for a sample from the SEVIR dataset.
Figure 28: Qualitative forecast. Forecast steps in 10 min intervals for a sample from the SEVIR dataset.
Figure 29: Qualitative forecast. The zoom-in highlights high variability among ensemble members for a min forecast.
Figure 30: Qualitative comparison. Compared to FGN ( Alet et al., 2025 ) , CasCast ( Gong et al., 2024 ) , FlowCast ( Ribeiro & Pucer, 2026 ) , and FREUD ( Schusterbauer et al., 2026b ) JWS better captures local and global precipitation patterns. model trained by us.
Figure 31: Qualitative comparison. FlowCast ( Ribeiro & Pucer, 2026 ) and JWS capture localized high precipitation accurately. model trained by us.
Figure 32: Qualitative comparison. JWS produces sharp forecasts that align with the ground truth and remains more realistic over leadtime than FREUD and CasCast.
Figure 33: Qualitative comparison. For low precipitation JWS is able to capture dynamics more accurately than FGN ( Alet et al., 2025 ) , CasCast ( Gong et al., 2024 ) , FlowCast ( Ribeiro & Pucer, 2026 ) , or FREUD ( Schusterbauer et al., 2026b ) . model trained by us.
Figure 34: Qualitative comparison. JWS produces sharper and more intricate precipitation patterns than prior work. model trained by us.
Figure 35: Ensemble variance. We show five ensemble members for samples from the SEVIR dataset at with complex precipitation patterns. Variance maps highlight localized uncertainty in complex precipitation regions.
Figure 36: Qualitative Inference-time scaling. At JWS produce accurate global structure with 1 NFE, 2 NFE resolves patch artifacts, and 10 NFE is close to the 50 NFE setting, with 50 NFE adding more local details and variability.
Figure 37: Qualitative Inference-time scaling. JWS captures finer structures consistently with more function evaluations at lead time.
Figure 38: Qualitative inference-time scaling. Results at lead time show that forecasts become blurry with few-step inference, while global structure is captured in all settings, and more sampling steps recover details.