acceptodds
Under review as a conference paper at ICLR 2027

Just 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 faster inference. Code and models will be released upon acceptance.

Primary area:
generative models
Keywords:
weather forecasting, generative models, flow matching, diffusion, nowcasting

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_LKzcOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes JWS, a single-stage, end-to-end diffusion model for probabilistic precipitation nowcasting that operates directly in radar space with a hierarchical transformer backbone. The authors introduce two components. Masked Asynchronous Diffusion (MAD) is a timestep-sampling scheme that keeps context frames clean while assigning correlated, resolution-aware noise levels to the predicted frames. Simple Distributional Diffusion Models (sDDM) train the denoiser with a timestep-dependent interpolation between CRPS and squared error to enable few-step sampling. On SEVIR and MeteoNet, the authors report state-of-the-art CRPS and SSIM, improved calibration, and lower inference latency and fewer parameters than multi-stage latent baselines such as CasCast and FREUD.

Critical strengths
  1. Simple, well-motivated design. JWS is a single end-to-end pixel-space diffusion model built on a standard hierarchical transformer. It avoids separately trained compression or deterministic-prior stages, and with them the reconstruction uncertainty those stages introduce.
  2. Effective few-step generation. The simple distributional objective (Eq. 7) removes the tuned schedules of DDM/iDDM. Controlled comparisons on the same architecture (Fig. 6; Fig. 7, right) show clear gains over Flow Matching, Shortcut models and FGN at low NFE, and one-step JWS-B produces a 10-member ensemble in under a second.
  3. Strong probabilistic skill. JWS achieves the best CRPS and SSIM on both SEVIR and MeteoNet (Tables 4–5). The T/S/B variants are better calibrated than CasCast and FREUD, with lower RI and flatter rank histograms (Fig. 12).
  4. Parameter efficiency. The 9M-parameter JWS-T already attains lower CRPS than FREUD (521M) and CasCast (402M).
  5. Flexible conditioning. MAD supports variable context, zero-shot deterministic priors and zero-shot extrapolation to 36 frames (Figs. 10, 14, 19).
Critical weaknesses
  1. Unclear method presentation. Fig. 2 only contrasts the FREUD and JWS pipelines, repeating Fig. 1, and does not show how JWS works: context handling, MAD noise levels, particle expansion, ξ injection, the loss and inference. Key components are delegated to an anonymous concurrent submission (iDDM), and sDDM differs from iDDM only in the loss interpolation (Eq. 7). The paper does not specify in which space and with which estimator the CRPS term is computed. It also does not explain the role of ξ at inference, where particles are dropped.

  2. Accuracy and speed come from different configurations. Table 4 uses 50 NFE, while latency is compared at 20 NFE. The statement that JWS remains faster at 50 NFE seems to hold only for the smaller variants. JWS-L, which provides the best CRPS/SSIM, takes 2.41s at 20 NFE, or ≈6 s at 50 NFE if latency scales linearly, versus 3.37 s for CasCast. The 17× speed-up refers to JWS-T, and the best RI also comes from the small variants: JWS-L's RI is worse than FREUD's.

  3. Gains are limited to metrics aligned with the loss. Without CFG, all JWS variants trail CasCast and FlowCast in HSS and CSI (e.g., CSI 0.4183 for JWS-L vs. 0.4401 and 0.4651). CFG adds an unconditional pass per step, and its latency is not reported. The improvements are mainly in CRPS, which is also the training objective.

  4. Baselines are not evaluated under a common protocol. Most baseline numbers are quoted from three papers, although all methods are said to be compared with 10 members. Because the CRPS estimator in App. C.1 is not the fair one, CRPS depends on ensemble size. RI is recomputed while the other metrics are quoted. PreDiff is trained at 128², and different FREUD variants appear in Table 4, Fig. 9 and Fig. 13. In addition, JWS trains on the merged training and validation sets with exhaustive subsequence sampling.

  5. Compute is under-reported. JWS-L requires 156.8 inference TFLOPs, 2.6× CasCast (Fig. 13), and the conditions under which the baselines were timed are not described. The training GPU type and GPU-hours are not reported, and the FLOP estimate behind the 44× training-efficiency claim is not explained. JWS-L also needs an additional pretraining phase on 128² crops .

  6. Possible test-set model selection. The validation set is merged into training. The 512 samples used for ablations and for choosing the training length therefore appear to come from the test set. Moreover, temporal consistency is not evaluated reported.

Questions for authors
  1. Could you add a figure showing the full JWS pipeline: context frames, MAD noise levels, particle expansion, ξ injection, the loss and inference?
  2. How does sDDM differ from iDDM (Anonymous, 2026) beyond Eq. 7, and in which space and with which estimator is the CRPS term computed?
  3. What are the latency and FLOPs of each variant at the 50 NFE used in Table 4, including the CFG setting?
  4. Were all baselines evaluated with the same test set, ensemble size, NFE and training data? If not, could you re-evaluate them with your pipeline?
  5. What GPU type and how many GPU-hours were used for training, and how were the training FLOPs in Fig. 1 computed?
  6. Is the 512-sample subset used for ablations and for choosing the training length drawn from the test set?
  7. Could you report a temporal-consistency metric and explain the abrupt change at +50 min in Fig. 32? If the authors can satisfactorily address the questions above, I am willing to raise my score.
5 comments
AuthorsOpenReview ↗

Response 1/2

Dear Reviewer LKzc,

We thank you for highlighting our “simple, well-motivated design,” enabling few-step generation, as well as our “strong probabilistic skill” with JWS. Below, we answer your remaining questions.

W1 & Q1 & Q2 - Pipeline Details

Thank you for the suggestion. We added Fig. 13 (appendix), which shows the full JWS pipeline used during training and inference. At inference, no particles are expanded: each ensemble member is generated with its own initial noise and xi. Thus, xi is held fixed for all ODE steps.

We added details on the sDDM loss in section A, appendix. To summarize, we use the default biased ensemble estimator of CRPS during training. CRPS and MSE are calculated per pixel in velocity space. iDDM requires the design and tuning of two timestep-dependent schedules. sDDM replaces them with a fixedinterpolation between CRPS and MSE. Although simpler, it consistently performs better across all metrics (Fig. 6).

W2 & Q3 & W3 - Latency and FLOPs

Tab. 15 reports wall-clock time and inference FLOPs. We also added Fig. 28, showing JWS latency over NFE including CFG. JWS-L takes 6.03s without guidance (12.09 s with guidance, Fig. 28), which is comparable to CasCast at matched settings (6.15 s without and 11.64 s with CFG, fp32) and 2-8x faster than FREUD and FlowCast. The 17x speed-up refers to JWS-T relative to the default CasCast configuration.

W2 - JWS-L RI

We looked further into why the largest model is less calibrated than FREUD. Fig. 18b shows that calibration depends on the training stage. After crop pretraining, JWS-L reaches an RI of 0.092 at 150k steps, below FREUD (0.136). With further full-resolution training, RI rises. The deterioration with additional training is consistent with overfitting and increasing ensemble underdispersion (Fig. 19c).

W3 - Gains beyond loss alignment

The improvements of JWS are not limited to CRPS. JWS also leads on SSIM and RI (T/S/B). With guidance, it achieves the best HSS on SEVIR and, at high thresholds, the best CSI (Fig. 26).

Moreover, the gains cannot be explained solely by alignment between objective and evaluation metric.

  • First, training directly on CRPS does not yield the best evaluation CRPS. In Fig. 6, a model trained only with a CRPS loss achieves a worse evaluated CRPS at 50 NFE than one trained with plain squared error. sDDM achieves superior performance. Likewise, FGN, trained with a CRPS objective, does not reach JWS’s CRPS (Fig. 7, right).
  • Second, a substantial part of the CRPS improvement comes from components unrelated to the loss. The timestep-sampling, paradigm, and architecture ablations (Tables 1-3) are all trained with standard Flow Matching (MSE).
  • Lastly, the training and evaluation CRPS differ. The loss uses a per-pixel CRPS over three velocity particles at a single noise level, whereas the evaluated CRPS is computed over ten fully sampled forecasts.
AuthorsOpenReview ↗

Response 2/2

W4 & Q4 - Evaluation under a common protocol

Our evaluation closely follows CasCast. We use the same test split and metrics computed with CasCast’s own SEVIRSkillScore class (https://github.com/OpenEarthLab/CasCast/blob/master/utils/metrics.py). This aligns our task and metric implementation with prior work. In Tab. 4, we quote reported values to include baselines without public checkpoints and present each method with previously established performance.

The new Tab. 13 (appendix) additionally compares CasCast, FlowCast, FREUD, and JWS using our pipeline on the same held-out test windows, with a fixed ensemble size and NFE across models. Reproduced baseline numbers deviate from the reported ones in both directions, so there is no bias in favor of our method. JWS models achieve the lowest CRPS, the highest SSIM, the lowest RI, and with CFG the highest HSS and CSI. We added FREUD-L to the top Fig. 9 and Fig. 14.

Regarding training data, all JWS models are trained on the same data as FREUD (Schusterbauer et al., 2026).

W5 - Timing and FLOP calculation

We added details to the supplement (Sec. C.2). Tab. 7 reports GPU type (GH200), GPU counts, and training times. Table 17 breaks down training FLOPs per model, measured as described in Sec C.2. The 44x claim compares JWS-T with FREUD, which JWS-T outperforms in CRPS. Appendix C.2 now describes the timing and FLOP measurement setup. Crop pretraining is included in all reported compute, but direct full-resolution training is competitive with comparable compute (Fig. 18).

W6 & Q6 - Ablation dataset

All model-selection decisions were made on a single fixed subset of 512 randomly selected SEVIR test windows (4.2%). Full-test evaluation was performed only after these choices were finalized and did not inform model selection.

To assess performance on events excluded from model selection, we removed every SEVIR test event containing any of these 512 windows, retaining 10,674 windows (87.8%), and evaluated JWS, CasCast, FlowCast and FREUD on this held-out set with the same pipeline (Table 13, appendix). JWS performance remains stable relative to the full test set. JWS variants achieves the best score on every metric.

Moreover, MeteoNet (Table 5) provides an additional check that is independent of model selection. We applied the same recipe selected on SEVIR unchanged.

W6 & Q7 - Temporal consistency

Thank you for spotting the mistake in Fig. 32 (now Fig. 39). When assembling the figure, the +50 min panel was taken from a different forecast.

Although the sDDM loss is applied per pixel, spatio-temporal consistency is learned by the backbone, as with a pixel-wise Flow Matching objective. We now evaluate spatio-temporal consistency explicitly in Table 12. JWS variants achieve the best scores on five of six metrics. On the change ratio, JWS is slightly too static, yet achieves better results than FlowCast and FREUD-L.

Dear Reviewer LKzc,

Thank you again for your time and detailed feedback. Do our additional experiments and clarifications address your concerns? We are more than happy to clarify any remaining questions.

Reviewer_9EnbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the JWS, an end-to-end probabilistic precipitation nowcasting model that generates forecasts directly in radar space, without separately trained compression or deterministic forecasting components. Its main ingredients are Masked Asynchronous Diffusion, which preserves clean conditioning frames while sampling correlated framewise noise levels; a spatiotemporal dimensionality-aware timestep shift; and a simplified distributional objective that interpolates pixelwise CRPS and particlewise MSE. A hierarchical transformer and late particle expansion reduce the computational cost of distributional training. Experiments on SEVIR and MeteoNet report strong CRPS performance across model sizes, competitive few-step generation, and substantial inference savings for the smaller variants. The paper also includes extensive component ablations, calibration analyses, and an extended-horizon experiment that uses deterministic prior conditioning.

Critical strengths
  1. The paper demonstrates that a relatively simple radar-space model can achieve strong probabilistic forecasting performance without separately trained compression and reconstruction stages. The latency and FLOP analyses further support the efficiency of the smaller models. This combination of system simplicity, ensemble skill, and low inference cost is consequential for frequently updated nowcasting.

  2. The experiments examine timestep sampling, masking, architecture, training objective, inference steps, ensemble size, solver choice, particle count, and particle-expansion position. In particular, Figure 6 provides useful evidence for the interpolated objective: purely distributional training performs well with few steps but deteriorates at higher NFE, whereas sDDM offers a more favorable balance.

  3. The paper also reports trade-offs, including the effect of guidance on calibration, rather than presenting every design choice as uniformly beneficial. These analyses make the work useful beyond its final benchmark numbers.

Critical weaknesses
  1. Equation (7) replaces the multivariate energy score with pixelwise CRPS and MSE. While this improves marginal forecasting performance, the objective does not explicitly constrain spatial and temporal dependencies between predicted frames. Different forecast ensembles with identical pixelwise marginals but different spatiotemporal correlations can achieve the same loss. This limits the theoretical support for the claimed distributional coverage, which is particularly important for coherent precipitation nowcasting.

  2. Several key components build directly upon prior work, including masked conditioning, timestep shifting, hierarchical transformers, and late particle expansion. More importantly, Table 2 shows that MAD does not improve CRPS over RaMViD, despite modest improvements in CSI and RI. Similarly, Figure 6 shows relatively small differences between sDDM and iDDM, which already employs time-dependent interpolation between distributional and regression objectives. These results raise questions about the extent of the methodological advance beyond an effective combination of existing techniques.

Questions for authors
  1. On MeteoNet, classifier-free guidance improves CSI from 0.2959 to 0.3448 but worsens RI from 0.0773 to 0.5719. What explains this substantial calibration degradation, and how should guidance strength be selected when both predictive accuracy and uncertainty calibration matter?

  2. MAD is designed to support arbitrary clean conditioning frames. How does JWS perform with missing or irregularly spaced observations compared with RaMViD? This would help establish the practical benefit of asynchronous conditioning.

  3. Figure 10 evaluates zero-shot extrapolation using an EarthFormer deterministic prior. How does JWS perform without this additional model? This would help distinguish its intrinsic extrapolation capability from the benefit of external conditioning.

  4. Tables 4 and 5 report CSI and HSS averaged over precipitation thresholds. How does JWS perform specifically on rare, high-intensity precipitation events? Threshold-specific results would help assess its ability to capture extreme weather.

3 comments
AuthorsOpenReview ↗

Response (1/2)

Dear Reviewer 9Enb,

Thank you for acknowledging that our "simple radar-space model can achieve strong probabilistic forecasting performance” and that our “analyses make the work useful beyond its final benchmark numbers”.

We address your questions below.

W1 - spatio-temporal consistency

We agree that the summed pixelwise CRPS in Eq. 7 is a score for each pixel on its own, but not for the joint distribution.

Pixelwise supervision does not, however, imply independent prediction, as for example diffusion models are also only trained with a pixelwise supervision signal. Similar to diffusion, JWS processes each member’s full context jointly through attention, allowing dependencies to be represented within a single network evaluation. In multi-step sampling, each update further conditions on the entire current input. These mechanisms allow the model to represent coherent forecasts, but do not by themselves establish correct joint structure. We therefore evaluate spatial and temporal dependence directly using multivariate scores, aggregate distributions, and temporal diagnostics in Table 12 and Fig. 25 of the Appendix.

Fig. 25 (Appendix) shows that joint quality improves with more steps. With a single step, JWS-B already outperforms CasCast on both energy scores and on the CRPS of frame-to-frame differences, but trails FREUD on the energy scores and both baselines on the change ratio. At 10 NFE, JWS matches or outperforms CasCast and FREUD on most metrics. At 50 NFE, JWS-S/B/L achieve the best energy scores of all compared methods, and JWS-S/B the best patch-mean CRPS. JWS-T is on par, and the temporal metrics are comparable to the baselines.

These results provide empirical evidence that JWS captures spatial and temporal structure beyond pixelwise marginal skill.

W2 - key components

We agree that several building blocks have precedents, and we cite them: masked conditioning (RaMViD), resolution-dependent timestep shifting (SD3), hierarchical transformers (HDiT) and distributional diffusion models with deferred expansion (iDDM). Our contribution is not any single block. It is a) finding which adaptations are needed to make single-stage, pixel-space probabilistic nowcasting work, b) reducing them to a simple recipe that practitioners can adopt directly, and c) showing that this recipe beats much larger multi-stage systems at a fraction of the cost (9M parameter model vs ~500M models on MeteoNet and SEVIR).

Naive transfer does not work. Moving the pixel-space image recipe directly to radar sequences fails. For example, JiTs logit-normal sampling with noise scaling gives RI 0.8575 (Table 1). Using a uniform timestep distribution with spatio-temporal resolution shift improves CRPS from 0.0211 to 0.0203 and RI by eight times (0.1066). Applying Diffusion Forcing naively also breaks forecasting: CRPS is 0.0317 versus 0.0203 (Table 2).

MAD vs. RaMViD (Table 2). MAD generalizes both paradigms: σ=0\sigma=0 recovers RaMViD, and pU=1p_U=1 with independent uniform timesteps recovers Diffusion Forcing. Its purpose is to add asynchronous per-frame noise levels without sacrificing clean-context skill. In Table 2, MAD is on par in CRPS (0.0205 vs. 0.0203) and improves CSI (0.3545 vs. 0.3467). The gain is larger at B scale: in Table 9 (appendix), increasing the spread from σ=0\sigma=0 (RaMViD) to σ=0.1\sigma=0.1 improves CRPS from 0.0194 to 0.0190, HSS from 0.474 to 0.515 and CSI from 0.364 to 0.397. MAD further has the benefit of per-frame noise levels, which help when context is imperfect (see our answer to Q2).

sDDM vs. iDDM (Figure 6). iDDM requires the design and tuning of two timestep-dependent schedules, λ(t)\lambda(t) and β(t)\beta(t). sDDM replaces them with a fixed, hyperparameter-free interpolation between CRPS and MSE. Although simpler, it consistently performs better across all metrics (Figure 6). For scale, this gain is about as large as going from JWS-S to the 4x larger JWS-B (Table 4: CRPS 0.0178 to 0.0176, CSI 0.4116 to 0.4180), at no extra cost.

System-level impact. Prior diffusion nowcasting models train two or three interdependent stages with 105M-521M parameters. JWS is a single model.

  • On SEVIR, our 9M JWS-T beats CasCast (402M) and FREUD (521M) in CRPS, and our 24M JWS-S beats FlowCast (160M).
  • On MeteoNet, our smallest model, JWS-T, beats every baseline in CRPS.
  • JWS needs 44x less training compute than FREUD at comparable CRPS and gives more than 17x faster inference.
  • sDDM also makes few-step generation possible with competitive probabilistic skill. One-step JWS-B produces a 10-member ensemble in under a second and is competitive with CasCast and FREUD (Fig. 9). At 1 NFE it outperforms Flow Matching, Shortcut models and FGN (Fig. 7).

We believe that this reduction in complexity matters for AI4Science, where broad adoption depends on pipelines that are simple to train, reproduce and deploy.

AuthorsOpenReview ↗

Response (2/2)

Q1 - analysis of guidance

We have added an analysis of calibration and forecasting performance under guidance in the appendix (“Analysis of CFG”).

What explains the degradation? CFG concentrates the ensemble and shifts it towards higher precipitation (Fig. 29). As a result, the reduced ensemble spread might no longer cover the error, and the observation falls below all members more often. This is the overconfident, wet-biased behavior that the rank histogram and RI penalize. CSI rewards the same shift, because more pixels exceed the fixed thresholds and detections increase, but so do false alarms. The per-pixel ensemble distributions (Fig. 41) show both sides. This trade-off is not specific to JWS as CasCast and FREUD behave the same way under guidance.

How to choose guidance strength? Accuracy at fixed thresholds and calibration pull in opposite directions under cfg. We therefore use unguided sampling, and all our calibration claims refer to that setting. If both matter, weak guidance is a reasonable compromise (recovers most of categorical gain while probabilistic skill degrades only mildly), while even weak guidance can harm calibration. Stronger guidance worsens all metrics (except probability of detection).

Q2 - asynchronous conditioning

Both, RaMViD and MAD are trained with a random subset of clean frames at random positions, so both models support arbitrary compositions of clean conditioning frames. Accordingly they are tied for CRPS when context frames are removed at random:

Removed context framesRaMViDMAD
4 random0.01930.0195
9 random0.01960.0197
4 most recent0.02730.0262

What MAD adds over RaMViD is per-frame noise levels, which matter when observations are available but imperfect. We tested two kinds of imperfect observations: reduced resolution and sensor noise. Both models can use the frames as clean, treat them as missing, or replace them; only MAD can also condition on them at an intermediate noise level, since RaMViD has never seen partially noised context and collapses when conditioned this way.

Reduced resolution. As a first experiment, we downsampled the last four context frames by average pooling and report each models best strategy:

Downsampling of last 4 framesRaMViD CRPSMAD CRPS
4x0.02140.0209
8x0.02350.0227
16x0.02700.0259

The differences are significant under a paired bootstrap over events: −0.0005 at 4x (95% interval [−0.0008, −0.0003]), −0.0009 at 8x ([−0.0012, −0.0005]) and −0.0010 at 16x ([−0.0014, −0.0007]).

Sensor noise. As as second experiment, we added Gaussian noise with standard deviation σ\sigma to the last four context frames, as a simple model of sensor noise (metric is CRPS, bold best per model):

σ\sigmaRaMViD cleanRaMViD discardMAD cleanMAD discardMAD intermediate noise level
0.250.02780.02730.02890.02620.0234
0.50.03500.02730.03670.02620.0233
1.00.04010.02730.04240.02620.0238

Conditioning at the intermediate noise level is better than both treating the frames as clean and discarding them. RaMViD’s best option is to discard the frames, and MAD improves on it by 0.0039, 0.0040 and 0.0034 CRPS (all 95% intervals exclude zero, e.g. [−0.0043, −0.0035] at σ=0.25\sigma = 0.25). Even very noisy frames σ=1\sigma=1 thus carry information that MAD can use.

Hence, both methods support missing and irregularly spaced clean observations, MAD is more robust when recent frames are missing or spatially degraded, and can additionally exploit noisy observations.

Q3 - long SEVIR

Thank you for this suggestion. We have added an evaluation without a deterministic prior to Fig. 10b) and compare it to other models without a deterministic prior (FREUD and FlowCast). Without any additional forecasting model it achieves competitive CRPS and CSI, while providing better calibration (RI=0.183, vs 0.298 for FREUD and 0.443 for FlowCast).

Q4 - CSI/HSS over precipitation thresholds

We have added per-threshold results for all JWS variants and baselines to the appendix (Fig. 26). Across thresholds, JWS is competitive in CSI, HSS, and POD, and has the lowest FAR of all models. We also report a F-beta score that combines POD and FAR (see lines 1378ff. in appendix). Figure 27 compares precision-focused, balanced, and recall-focused settings over thresholds. With guidance, JWS improves POD, CSI, and HSS at high thresholds, but at the cost of higher FAR, which still remains below CasCasts at better calibration.

We further evaluate performance on the high-rain events in the SEVIR test set, restricted to the 90th percentile (Table 14). JWS maintains its advantage in distributional quality and calibration and performs comparably with baselines on SSIM, HSS, & CSI.

Dear Reviewer 9Enb,

Thank you again for your time and constructive suggestions. Do the additional experiments with noisy and degraded observations, together with our discussion of the contributions relative to prior work, help address your concern about the practical and methodological advances? Please let us know if this or any of your other concerns would benefit from further clarification.

Reviewer_DNpkOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Just Weather Scoring (JWS), an end-to-end framework for probabilistic precipitation nowcasting directly in radar space. Building on clean-data prediction with velocity-space supervision, it proposes Masked Asynchronous Diffusion (MAD), spatiotemporal dimensionality-aware timestep shifting, and a simple distributional objective (sDDM) that interpolates between pixelwise CRPS and MSE. Shared early network computations reduce the overhead of multi-particle training, while the model supports few-step inference without separately trained compression or deterministic forecasting components. Experiments on SEVIR and MeteoNet demonstrate strong probabilistic forecasting performance and substantial efficiency gains. Extensive ablations examine the contributions of the sampling strategies, objectives, architecture, and training configurations.

Critical strengths
  1. Useful methodological contributions supported by systematic component analyses.
    The clearest contribution is sDDM's prescribed, timestep-dependent interpolation between pixelwise CRPS and MSE, which replaces iDDM's tuned loss schedules without introducing additional loss-schedule hyperparameters. Figure 6 supports its effectiveness at matched sampling budgets. MAD combines clean observed context with correlated, asynchronous noise levels, while timestep shifting accounts for the full spatiotemporal dimensionality. The extensive component analyses substantiate these design choices and make the methodological contribution a central reason for my positive assessment.

  2. Strong forecasting performance with substantial practical efficiency gains.
    JWS demonstrates strong CRPS and SSIM results on both benchmarks, alongside strong calibration for several unguided configurations. Its 9M-parameter variant remains competitive with substantially larger models. The single-model pipeline avoids separately trained compression and deterministic forecasting components, and few-step generation substantially reduces inference latency in the reported comparisons. This combination of forecasting quality and computational efficiency is particularly valuable for nowcasting, where predictions must be updated frequently under practical resource constraints.

Critical weaknesses

None

Questions for authors
  1. Contribution of timestep shifting -- additional validation.
    Table 1 compares spatial and spatiotemporal shifts but does not include an unshifted uniform baseline. Could you report this baseline with r(τ~t)=1r(\tilde\tau^t)=1, retaining σ=0\sigma=0 and the other settings used in Table 1? This would isolate the contribution of Eq. (4), complement the correlation ablations in Tables 2 and 8, and clarify the empirical importance of the proposed shifting scheme.

  2. Performance on intense events -- additional analysis.
    Could you provide CSI at individual thresholds, particularly for SEVIR? Although JWS achieves strong probabilistic scores, JWS-L with CFG does not surpass FlowCast in threshold-averaged SEVIR CSI. Threshold-specific results, together with probability of detection and false alarm ratio at higher thresholds, would clarify whether the method misses intense events or produces excessive precipitation predictions. This distinction matters to my assessment of its usefulness for precipitation nowcasting.

  3. Guidance and calibration -- clarification.
    On MeteoNet, CFG increases JWS-L's CSI from 0.2959 to 0.3448 and HSS from 0.4186 to 0.4816, while RI worsens from 0.0773 to 0.5719 and CRPS increases from 0.0138 to 0.0141. Could you clarify the implications of this trade-off for the paper's calibration claims and explain how practitioners should choose a configuration when calibrated probabilistic forecasts are the priority? This would help establish the scope of the reported improvements.

  4. Large-model training and the single-stage claim -- clarification.
    The reported JWS-L results use pretraining on smaller inputs followed by full-resolution fine-tuning. Since single-stage training is one of the major contributions of the paper, I think the dependency on pretraining should be discussed to clarify whether this two-stage training (pretraining then fine-tuning) is necessary. Figure 17 shows that this curriculum is beneficial, but can direct full-resolution training achieve comparable performance under a matched total computational budget? Please distinguish the absence of separately trained model components from any curriculum required to obtain the reported large-model performance.

3 comments

Dear Reviewer DNpk,

Thank you for appreciating our contributions, which are “supported by systematic component analysis” as well as our “strong forecasting performance” accompanied by “efficiency gains”. We answer your remaining questions below.

Q1 - Uniform timestep shift

Thank you for this suggestion. We added the unshifted uniform baseline to Table 1. Relative to this baseline, spatiotemporal shifting (Eq. 4) reduces CRPS (4.2%), increases CSI (7.8%), and reduces RI (12.7%).

Q2 - Performance on intense events

We added per-threshold CSI, HSS, probability of detection (POD), and false alarm ratio (FAR) on SEVIR to the supplementary material (Fig. 26), together with an F-beta analysis that weights misses and false alarms differently (Fig. 27). At low thresholds, unguided JWS performs on par with prior methods. At higher thresholds (beyond 133), unguided JWS trades detection against false alarms differently from the baselines. It achieves the lowest FAR of all methods at every threshold, but also a lower POD, which explains its lower threshold-averaged CSI. Prior methods, particularly CasCast, reach higher detection at the cost of substantially more false alarms. The F-beta scores confirm this: in the precision-weighted setting (F0.5), JWS-L performs on par with FlowCast and outperforms FREUD and CasCast. For thresholds beyond 133, the baselines lead in the balanced (F1) and recall-weighted (F2) settings. With CFG, JWS-L improves POD, CSI, and HSS at high thresholds, achieving the highest CSI and HSS at thresholds above 160 and the best F1 and F2. The cost is a higher FAR, which remains below CasCast’s, and worse calibration (Fig. 29).

Finally, Table 14 reports results for test windows with mean VIL exceeding the 90th percentile of the SEVIR test set. On this heavy-precipitation subset, unguided JWS retains the best CRPS (0.0365 vs. 0.0373 for FlowCast) and, for T/S/B-scale, the best RI (down to 0.0638 vs. 0.1056 for FREUD).

Q3 - Guidance & calibration

We have added an analysis of the effect of guidance in the appendix. For our main experiments we use unguided results. Guided results are reported separately in grey (Tables 4 and 5) to show the additional categorical skill and its associated trade-offs (e.g. worse RI/CRPS). These results answer the following questions:

  • Why does guidance hurt calibration? CFG sharpens the predictive distribution and shifts it towards heavier precipitation. On SEVIR, increasing the guidance scale raises precipitation intensity (Fig. 41) and detection rates, but also false alarms, and consistently degrades RI for all methods (Fig. 29).
  • Scope of the improvements? CRPS remains competitive despite worse calibration (RI), the small change in CRPS therefore does not imply preserved calibration. On MeteoNet, JWS-Ls CRPS increases only from 0.0138 to 0.0141, which remains well below the baselines (CasCast 0.0180, FREUD 0.0193). On SEVIR, JWS-L at CFG 1.5 still achieves better CRPS than all baselines. Calibration as measured by RI, however, degrades substantially, most strongly for JWS-L on MeteoNet.
  • How to choose a configuration? When calibrated probabilistic forecasts are the priority, we recommend sampling without guidance. When categorical skill at fixed thresholds matters, moderate guidance improves CSI/HSS. Weaker guidance (~1.25) offers a compromise with nearly unchanged CRPS (Fig. 29)

Q4 - Large-model training

Thank you for highlighting this. By “single-stage”, we mean that JWS is one network trained end-to-end with one objective, without separately trained compression, decoding, or deterministic forecasting components. Only JWS-L uses a training curriculum (100k steps on random 128px crops, then 100k steps at 384px). JWS-T/S/B are trained in a single phase at full resolution and already outperform or match baselines in CRPS on SEVIR. Our main results therefore do not depend on the curriculum.

Direct full-resolution training of JWS-L at a comparable budget is competitive (Fig. 18 & table below). The best evaluated from-scratch checkpoint (100k steps) uses about 10% less compute than the curriculum and reaches CRPS 0.0183 vs. 0.0175 and CSI 0.415 vs. 0.418, with comparable calibration. Longer training does not help as the model overfits (CRPS 0.0188 at 200k steps, 1.8x the curriculum compute). The curriculum therefore has two benefits. First, it mitigates overfitting, and second, it reduces training cost, because the first 100k steps run at lower resolution.

We therefore use the curriculum for JWS-L. Fig. 19 further suggests that the curriculum might compensate for limited data diversity: with sufficient training data, direct JWS-L training is stable.

Training JWS-LStepsCRPS ↓CSI ↑RI ↓SSR → 1
Direct 384px100k0.01830.4150.1830.818
Direct 384px200k0.01880.4040.1680.790
Curriculum (128px -> 384px)100k + 100k0.01750.4180.1910.844

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my questions. While the responses clarify my original concerns, they also reveal two limitations that somewhat temper the strengths I initially highlighted.

First, regarding practical forecasting gains, the newly reported per-threshold results suggest that the strong overall probabilistic forecasting performance does not necessarily translate into superior detection of intense precipitation. In the unguided setting, JWS achieves lower false alarm ratios at the expense of detection rates, while competing methods perform better under balanced and recall-weighted F-beta metrics at higher thresholds. Although classifier-free guidance improves detection performance, it introduces a trade-off with calibration. Given the practical importance of detecting intense precipitation events in nowcasting, these findings somewhat weaken the practical significance of the reported improvements.

Second, regarding training simplicity and scalability, the best-performing JWS-L model relies on a curriculum involving smaller spatial crops before full-resolution training. Direct full-resolution training produces worse results and exhibits overfitting with additional training. While I acknowledge that this does not contradict the single-stage nature of the proposed framework, and that smaller JWS variants perform well without such a curriculum, it suggests that robust scaling to larger models is less straightforward than the paper's emphasis on training simplicity might imply.

Overall, I appreciate the authors' thorough responses and additional experimental evidence, which have satisfactorily addressed my questions. I also acknowledge that the revised manuscript now discusses these limitations.

Dear Reviewer DNpk,

Thank you for your response and for giving us the opportunity to address your remaining reservations. We are happy that we were able to satisfactorily clarify your previous questions.

Q2.1: Practical forecasting gains

We think there might be a misunderstanding, and our previous response (and Fig. 26) might not have communicated our results clearly enough. We added Fig. 30 to the paper and hope this makes it clearer.

While calibration and detection do present a trade-off (Fig. 29), JWS achieves strong performance on both axes (Fig. 30). This also holds at high precipitation thresholds (Fig. 26, Fig. 27, Fig. 30). We achieve superior calibration performance (CRPS, RI) and superior detection (CSI) with a guidance scale of 1.5.

Hence, despite simplifying the overall pipeline (a single model) and reducing model size (94M vs 402M/521M), we score best on both calibration and detection. For our main paper, however, we put emphasis on calibration; hence, we use an unguided model as the default setting.

Additionally, Table 14 shows that unguided JWS maintains superior CRPS and RI while achieving comparable SSIM, HSS, and CSI when evaluating only high-precipitation events. Therefore, JWS with moderate guidance (≤ 1.5) provides strong practical forecasting utility when detecting extreme events and producing calibrated probabilistic forecasts are equally important.

Q2.2: Training simplicity and scalability

While we need data augmentation to achieve superior performance with JWS-L, we argue that this is not a limitation of our proposed method, but rather of the availability of public data. Scaling usually requires scaling all axes, and we show in Fig. 19 that our method indeed scales cleanly when data is not the limiting factor.

However, removing JWS-L from Tab. 4 (and also Tab. 5) leaves the main conclusions and contributions unchanged: JWS-B is trained without a curriculum and yields the best CRPS and SSIM, superior calibration, and competitive HSS and CSI. As described above, with guidance, we also get superior HSS and CSI (Fig. 26). The results presented in our paper therefore do not hinge on the inclusion of JWS-L. Even our smallest 9M-parameter model achieves competitive performance.

We hope these responses could alleviate any remaining reservations, and thanks again for your time.

Reviewer_WRymOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

JWS performs probabilistic precipitation nowcasting directly in radar space. It combines masked asynchronous diffusion, dimensionality-aware timestep sampling, and a loss that interpolates between pixelwise CRPS and particlewise MSE. The model supports few-step generation without a separately trained compression model. Experiments on SEVIR and MeteoNet examine forecast quality, calibration, and computational cost.

Critical strengths
  1. The paper makes a convincing practical case for direct radar-space modeling. Tables 4–5 show strong CRPS and SSIM results across model sizes, while Figure 9 demonstrates a useful accuracy–latency trade-off. The small models are particularly appealing: competitive forecasts do not require a large model or a separate compression pipeline.
  2. The experiments go well beyond the final benchmark comparison. Section 4.1 separates timestep sampling, conditioning, architecture, and training objective, and Figure 7 includes matched comparisons with other few-step methods. The objective ablation is informative: it shows why combining distributional supervision with regression works more consistently across sampling budgets than either alone.
Critical weaknesses
  1. The evidence for probabilistic quality is mainly at the level of individual pixels. Equation (7) supervises conditional marginals, unlike the multivariate energy score introduced in Equation (6). Good CRPS and pooled rank histograms therefore do not establish that the ensemble captures spatial and temporal dependence correctly. The qualitative ensembles are useful, but this remains an important gap for the one- and few-step results, where getting the joint structure of a storm right matters alongside its pixelwise distribution.
  2. The separation between model selection and final evaluation is unclear. Appendix C.3 combines validation data with training, while C.2 uses 512 samples for the ablations that motivate the final recipe. I could not determine which split supplies those samples or which data inform hyperparameter and checkpoint choices. This needs clarification before the final test gains can be interpreted as an independent evaluation of the selected recipe.
Questions for authors
  1. Do one- and few-step forecasts retain their advantage when evaluating the predictive distribution of a regional or temporal aggregate of the radar field, rather than individual pixels? A comparison with the 50-step model would help establish what is preserved as sampling is shortened.
  2. Which data were used to select hyperparameters and checkpoints, and do the 512 ablation samples overlap with the final test set? If they informed model selection, how do the results look on events excluded from that process?
2 comments

Dear Reviewer WRym,

Thank you for acknowledging the “convincing practical case for direct radar-space modeling” our paper presents, as well as our experiments, which “go well beyond the final benchmark comparison”.

We address your questions below:

W1 & Q1 - spatio-temporal consistency

We agree that getting the joint structure of the storm right matters alongside its pixelwise distribution.

Pixelwise supervision does not, however, imply independent prediction, as for example diffusion models are also only trained with a pixelwise supervision signal. To further asses the spatial and temporal dependence directly we use multivariate scores, aggregate distributions, and temporal diagnostics and report results in Table 12 and Fig. 25 of the Appendix. The results indicate that distributional alignment extends beyond pixelwise marginals to spatio-temporal dependence.

Regarding few-step sampling, Fig. 25 (appendix) reports the energy scores and temporal metrics for JWS-B at different NFE. Joint quality improves monotonically with more steps. With a single step, JWS-B already outperforms CasCast on both energy scores and on the CRPS of frame-to-frame differences, but trails FREUD on the energy scores and both baselines on the change ratio, i.e., one-step forecasts slightly underrepresent temporal change. At 10 NFE, JWS matches or outperforms CasCast and FREUD on all metrics except on change ratio.

We further include an additional qualitative sample in Fig. 35, which shows that JWS captures the complex spatio-temporal evolution of a known hurricane event (Hurricane Dorian).

W2 & Q2 - model selection

Following FREUD (Schusterbauer et al., 2026), we merge the official validation split into training. All model-selection decisions (hyperparameters, checkpoints, training duration, and guidance scale) were made on a single fixed subset of 512 randomly selected SEVIR test windows (4.2%). Full-test evaluation was performed only after these choices were finalized and did not inform model selection.

To assess performance on events excluded from model selection, we removed every SEVIR test event containing any of these 512 windows, retaining 10,674 windows (87.8%), and evaluated JWS, CasCast, FlowCast and FREUD on this held-out set with the same pipeline (Table 13, appendix). JWS performance remains stable relative to the full test set, as CRPS, HSS and CSI of all JWS models change below 1% and SSIM by at most 1.2%, while RI improves. Compared to the baselines, JWS achieves the best score on every metric: CRPS (0.0175 for JWS-B/L vs. 0.0177 for FlowCast), SSIM (JWS-S), RI (JWS-T), and, with CFG, HSS and CSI (0.4570 vs. 0.4536 for FlowCast).

MeteoNet (Table 5) provides an additional check that is independent of model selection. We applied the same recipe selected on SEVIR unchanged. No MeteoNet test data informed any design decision. The reweighting experiment in Supplementary Table 11 was a post-hoc diagnostic and did not alter the recipe. JWS achieves the best reported CRPS and SSIM on MeteoNet and, with classifier-free guidance, the best reported HSS and CSI among the compared methods.

Dear Reviewer WRym,

Thank you again for your time and thoughtful feedback. Do the added spatiotemporal evaluations (Table 12 and Fig. 25) and the evaluation in Table 13 address your two main concerns? Please let us know if any questions remain unanswered or if further clarification would be helpful.

Comments on the paper

2 comments

Dear reviewers,

Thank you for your constructive feedback.

We are glad that you recognize our strong CRPS and SSIM results (WRym, DNpk, 9Enb, LKzc), strong calibration (DNpk, LKzc), and “simple, well-motivated design” (LKzc) that avoids separately trained compression and deterministic stages (WRym, DNpk, 9Enb, LKzc). We particularly appreciate your recognition of our “effective” (LKzc) and “competitive” (9Enb) few-step generation, “particularly appealing” small models (WRym), and “substantial practical efficiency gains” (DNpk), which you consider “particularly valuable for nowcasting” (DNpk) and “consequential for frequently updated nowcasting” (9Enb).

You also value the methodological insights our work provides (WRym, DNpk, LKzc), as well as its flexible conditioning capabilities (LKzc). In particular, you highlight that our experiments “go well beyond the final benchmark comparison” (WRym), with “extensive” (DNpk, 9Enb) and “systematic” (DNpk) component analyses, making “the work useful beyond its final benchmark numbers” (9Enb).

We highlight updates to the paper in blue\color{blue}{\text{blue}} and address each reviewer’s comments individually below.

Dear AC and Reviewers,

Thank you for the time you invested in reviewing our paper.

We are happy that you pointed out the following strengths of our submission:

  • Substantial efficiency gains (DNpk, LKzc, 9Enb) with a competitive 9M parameter model and strong few-step generation performance
  • Strong probabilistic forecasting performance (DNpK, 9Enb, LKzc) without separately trained compression and forecasting components (9Enb, DNpk, LKzc)
  • Extensive experiments and ablations (WRym, DNpk, LKzc) that “go well beyond final benchmark comparison” (WRym), with extensive and systematic component analyses (9Enb, DNpk)
  • A simple model (9Enb, LKzc) with useful methodological contributions (DNpk) and a convincing practical case for direct radar-space modeling (WRym)

In addition to expanding details on our method, we added the following experiments in response to your questions:

  • Inclusion of a uniform, unshifted timestep sampling baseline (Tab. 1)
  • Spatio-temporal consistency metrics (Tab. 12 & Fig. 25)
  • SEVIR-long extrapolation results without a prior (Fig. 10)
  • Extended JWS-L calibration per data size (Fig. 18 & Fig. 19)
  • Per-threshold CSI, HSS, PoD and FAR (Fig. 26 & Fig. 27)
  • Performance on high-precipitation data (Tab. 14)
  • Extended latency analysis (Fig. 28 & Tab 15)
  • Extended analysis of CFG (Fig. 29)

We hope that these results and our clarifications address all points raised. We have integrated these suggestions and clarifications into the main paper and appendix.

Thank you for helping us improve our paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.