acceptodds
Under review as a conference paper at ICLR 2027

From Continuous Fields to Discrete Tokens: Probabilistic Precipitation Nowcasting with Controlled-Path Diffusion

Abstract

Precipitation nowcasting requires accurate forecasts of uncertain rainfall evolution at low computational cost. Diffusion models generate diverse forecasts, but their training and iterative sampling can limit practical use. We introduce Discrete Probabilistic Precipitation Nowcasting (DPPN), to our knowledge the first discrete diffusion framework for probabilistic precipitation nowcasting. DPPN encodes continuous radar fields into compact discrete tokens that form a learned vocabulary of precipitation patterns. By modeling their conditional evolution, DPPN offers a perspective on nowcasting as learning a weather language. DPPN connects continuous Gaussian relaxations with discrete token states to reduce training variance and improve computational efficiency. The framework supports two forecasting formulations: direct generation conditioned on historical observations and cascaded generation conditioned on a deterministic prediction. Both generate distributions over future tokens, with supervision on decoded precipitation fields guiding forecast reconstruction. Experiments on SEVIR demonstrate competitive forecasting performance relative to continuous diffusion and flow-matching baselines, with fewer trainable parameters and lower reported wall-clock costs under the evaluated implementations. The generated results highlight discrete diffusion as an efficient approach to probabilistic nowcasting and demonstrate its capacity for controllable generation through diffusion path design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
discrete diffusion, probabilistic precipitation nowcasting, radar tokenization, spatiotemporal generative modeling, diffusion path design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Cpp3OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors of this paper used VQ-VAE to discretize radar echoes into tokens, which were then fed into a diffusion layer to achieve probabilistic precipitation nowcasting. On SEVIR, the pooling scale mCSI (Pool16) reached 0.6348 with a 1-hour lead time. Furthermore, it can be trained on a single 32GB GPU.

Critical strengths

The advantages are:

  1. 128×128 radar frames are frozen and compressed into a 16×16 token grid (codebook 512) using VQ-VAE, with diffusion occurring on 256 discrete states per frame, rather than a dense, continuous latent state.

2.The diffusion path itself is controllable: a perplexed statistical construct is used to drive the erosion kernel, and combined with a precisely derived reverse rate and path-following sampler, resulting in significant benefits under low step budget.

3.Leading in pooling-scale event detection capabilities

Critical weaknesses

Weaknesses:

  1. Limited dataset: Experiments were only conducted on one dataset, and the method's generalization to other radar systems/climate zones was not verified. Incidentally, CasCast at least conducted experiments on SEVIR and TST-BRAZIL.

  2. Most metrics were not tested: In Table 1, 10 out of 13 methods had empty SSIM/CRPS/HSS/mCSI@60min parameters.

  3. Sensitivity experiments were not conducted for longer forecast time domains such as 90 min, 2, etc., or different historical lengths or frame intervals; however, the advantage of longer lead times is greater, but I didn't see how long this prediction duration actually was.

Questions for authors

See Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer Cpp3

Thank you for the detailed and constructive review. In response to the reviewer's concerns, we have substantially expanded and clarified the experimental evaluation in three aspects:

Cross-dataset evaluation. We add experiments on MeteoNet to evaluate the generalization of DPPN beyond SEVIR. DPPN-Cas achieves the best CSI-M and CSI-47 performance among the compared methods under both ensemble-mean and percentile-based fusion.

Clarification of baseline metrics and evaluation availability. We reorganize the original Table~1 into two tables to distinguish published benchmark results from methods for which suitable checkpoints and implementations enable our complete evaluation. This removes the ambiguity of the ``--'' entries and makes the source and scope of each reported metric explicit.

Longer-horizon evaluation. We clarify that the primary SEVIR setting uses seven historical frames to predict six future frames at 10-min intervals, corresponding to a 60-min forecast horizon. We further extend the trained model autoregressively to 120 min without retraining and report performance from 70 to 120 min. The model retains a pixel-wise mCSI of 0.27 at 90 min and 0.23 at 120 min and a Pool16 mCSI of 0.34 at 90 min and 0.30 at 120 min. We also explicitly acknowledge that systematic sensitivity to observation-window length, frame interval, and explicit longer-horizon training remains outside the scope of the present study.

Detailed responses are provided below.

AuthorsOpenReview ↗

Response to Reviewer Cpp3

Evaluation on MeteoNet

To evaluate DPPN on an additional radar dataset, we trained the tokenizer and DPPN-cas on MeteoNet, a reflectivity dataset from Météo-France. The system uses 12 observed frames to forecast the next 12 frames on a 400 × 400 grid at 5-minute intervals, for a 60-minute forecast horizon. (https://meteofrance.github.io/meteonet/english/data/rain-radar/)

On the test set, we obtain the final forecast using two ensemble-fusion strategies. The first is the conventional ensemble mean, computed by averaging the eight ensemble members pixel-wise in image space. The second is our percentile-based pixel-wise fusion described in Section 4.1. Specifically, at each pixel, we compute the median (P50) and the 60th percentile (P60) across the eight ensemble members. We select P60 when its intensity is at least 35 and P50 otherwise. CSI-M is the mean CSI over thresholds of 19, 28, 35, 40, and 47 dBZ, while CSI-47 evaluates detection specifically at 47 dBZ. For Pool4 and Pool16, non-overlapping block max pooling is applied to both the forecast and observation before computing the CSI metrics.

ModelCSI-M Pool1CSI-M Pool4CSI-M Pool16CSI-47 Pool1CSI-47 Pool4CSI-47 Pool16CRPS ↓
ConvLSTM [NeurIPS 15]0.30080.30500.34650.09820.10910.15880.0218
PredRNN [NeurIPS 17]0.29140.30030.34020.08230.09900.14620.0214
PhyDNet [CVPR 20]0.31200.31240.33560.11060.11570.14820.0216
SimVP [CVPR 22]0.30170.31430.35770.09970.11340.15990.0218
LDM [CVPR 22]0.21310.21910.23690.03590.04070.05520.0209
EarthFormer [NeurIPS 22]0.28310.28550.31540.07870.08720.12080.0224
PreDiff [NeurIPS 23]0.25460.26680.29350.04900.05940.08670.0197
NowcastNet [Nature 23]0.29550.32320.37340.12360.15210.21150.0277
CasCast (EarthFormer) [ICML 24]0.31560.36500.44200.12040.15630.23570.0180
DPPN (Earthformer)-mean0.3323‾\underline{0.3323}0.3834‾\underline{0.3834}0.41760.1290‾\underline{0.1290}0.19400.2657‾\underline{0.2657}0.0165
DPPN (Earthformer)-fusion0.33690.38850.46780.14140.1937‾\underline{0.1937}0.31140.0165

For both ensemble aggregation strategies, DPPN-Cas achieves the best performance among the compared methods in both CSI-M and CSI-47. Using the same Earthformer deterministic forecast, CasCast obtains CSI-M and CSI-47 scores of 0.4420 and 0.2357, respectively. FREUD+LSM-L [CVPR 2026] also reports results on MeteoNet, achieving an mCSI of 0.1417 and CRPS of 0.0193 under its specified data split. However, as its code and checkpoints are not available to us, we are unable to reproduce the evaluation under either our or their experimental setting for a direct comparison.

AuthorsOpenReview ↗

Response to Reviewer Cpp3

Missing metrics in Table 1

Thank you for pointing this out. We agree that the original Table 1 did not make the source and availability of the baseline results sufficiently clear. Our original intention was to provide a broad comparison with existing SEVIR forecasting methods in a single table. However, the availability of checkpoints and evaluation outputs differs across baselines. For several methods, including ConvLSTM and PredRNN, matching SEVIR checkpoints are not available to us. We therefore report their published CRPS and CSI results from the PreDiff benchmark (Table 2), but cannot consistently evaluate additional metrics such as SSIM, HSS, and 60-minute mCSI.

In the original Table 1, these unavailable evaluations were indicated by “–”. We agree that this notation can be misinterpreted as missing experiments rather than unavailable evaluations. PreDiff provides an official checkpoint, but reproducing the full ensemble-based evaluation is computationally expensive: one complete SEVIR test-set ensemble evaluation takes approximately 110 hours under our current setup. This computational cost also reflects one of the practical motivations for DPPN, which performs discrete diffusion on compact radar-token grids to reduce the cost of probabilistic forecasting.

To avoid ambiguity, we have reorganized the original Table 1 into two separate tables according to evaluation availability. The first table reports the metrics consistently available from the published benchmark, namely CRPS and aggregate mCSI at Pool1, Pool4, and Pool16. The second table contains methods for which suitable checkpoints and implementations are available to us, allowing a more complete and consistent evaluation including SSIM, CRPS, HSS, aggregate mCSI, and 60-minute mCSI. This separation avoids using “–” for metrics that cannot be evaluated and makes the source and scope of each comparison explicit.

Revised Table 1(a): SEVIR forecast quality using published benchmark results

ModelSize (M)CRPS ↓mCSI Pool1 ↑mCSI Pool4 ↑mCSI Pool16 ↑
Persistence—0.05260.26130.37020.4690
UNet16.60.03530.35930.40980.4805
ConvLSTM14.00.03320.41850.44520.5135
PredRNN46.60.03060.40800.44970.5005
PhyDNet13.70.03190.39400.43790.4854
E3D-LSTM35.60.02970.40380.44920.4961
Rainformer184.00.03570.36610.42320.4738
DGMR71.50.04350.26750.34310.4832
VideoGPT99.60.03810.36530.43490.5798
LDM438.60.02800.35800.40220.5522
PreDiff220.50.02460.41000.46240.6244

The first table therefore provides the broad comparison supported by the metrics available for these published baselines, without using “–” to represent metrics that cannot be consistently evaluated.

Revised Table 1(b): Complete SEVIR evaluation for models with suitable checkpoints and implementations

ModelSize (M)SSIM ↑CRPS ↓HSS ↑mCSI Pool1 ↑mCSI Pool4 ↑mCSI Pool16 ↑mCSI at 60 min Pool1 ↑mCSI at 60 min Pool4 ↑mCSI at 60 min Pool16 ↑
Earthformer15.10.76320.03040.53900.41630.43090.46720.29740.30340.3270
CasCast309.40.75870.02100.55380.40140.48850.62800.31370.35680.4554
FREUD+LSM-L473.00.73730.02010.52760.41410.46350.53950.30030.33220.3935
FlowCast230.40.78660.01860.56070.44170.48120.56090.32760.35720.4332
DPPN-direct92.60.83480.02030.53510.41140.46830.60950.31650.36710.5295
DPPN-Cas102.00.83310.02160.54660.41860.50490.63480.32880.37830.4742

The second table reports the complete set of metrics only for methods for which suitable checkpoints and implementations are available to us, allowing these metrics to be evaluated consistently. HSS averages the six thresholds at native resolution. In the revised manuscript, bold and underline indicate the best and second-best available results within each corresponding comparison table. We believe this restructuring makes the comparison more transparent: the absence of SSIM, HSS, or 60-minute mCSI for methods in Table 1(a) reflects evaluation availability rather than omitted experiments, while Table 1(b) provides the complete metric comparison where consistent evaluation is possible.

AuthorsOpenReview ↗

Response to Reviewer Cpp3

Forecast horizon and sensitivity

Thank you for pointing this out. We agree that the original manuscript did not make the forecast horizon sufficiently explicit and did not investigate sensitivity to longer forecast horizons, different observation-window lengths, or frame intervals.

The main experiments in the submitted manuscript use seven historical radar frames to predict the subsequent six frames at a 10-min interval, corresponding to a 60-min forecast horizon. To examine behavior beyond this training horizon, we have added an evaluation extending the forecasts to 120 min without retraining the model.

Specifically, the trained model first uses the seven observed frames to predict six future frames (10–60 min). We then construct a second seven-frame conditioning sequence from the last observed frame and these six predicted frames,

[x0,x^10,x^20,…,x^60],[x_{0},\hat x_{10},\hat x_{20},\ldots,\hat x_{60}],

and apply the same trained model again to predict [x^70,…,x^120][\hat x_{70},\ldots,\hat x_{120}].

Thus, we use the seven observed frames

[x−60,x−50,x−40,x−30,x−20,x−10,x0][x_{-60},x_{-50},x_{-40},x_{-30},x_{-20},x_{-10},x_{0}]

to predict

[x^10,x^20,…,x^60,x^70,x^80,…,x^120],[\hat x_{10},\hat x_{20},\ldots,\hat x_{60},\hat x_{70},\hat x_{80},\ldots,\hat x_{120}],

where xtx_{t} and x^t\hat x_{t} denote the observed and forecast frames at tt minutes relative to the forecast start. The 70–120 min forecasts are therefore obtained through an autoregressive extension of the original 60-min model, without additional training or fine-tuning. The first two forecast starts of each of the 4,053 four-hour SEVIR test events provide observations through 120 min, giving 8,106 samples for this evaluation.

Lead time (min)mCSI Pool1 EarthFormermCSI Pool1 DPPN-casmCSI Pool4 EarthFormermCSI Pool4 DPPN-casmCSI Pool16 EarthFormermCSI Pool16 DPPN-cas
700.29810.30300.31160.32570.34010.3656
800.27630.28580.28550.31180.30850.3579
900.25700.26920.26400.29470.28330.3432
1000.23960.25410.24520.27880.26190.3274
1100.22400.24050.22870.26430.24450.3141
1200.20890.22840.21320.25040.22750.2986

Skill decreases with lead time for both models. Beyond 60 minutes, most inputs to the second rollout step are forecasts from the first step, so first-hour errors carry into the second hour. Without retraining for the longer horizon, DPPN-cas scores above EarthFormer at every lead time and pooling scale, and the margin grows with lead time. At Pool16, DPPN-cas reaches 0.3432 at 90 minutes and 0.2986 at 120 minutes, against 0.2833 and 0.2275 for EarthFormer. The Pool16 gap widens from 0.0255 at 70 minutes to 0.0711 at 120 minutes, and the Pool1 gap from 0.0049 to 0.0195.

We have added these results to the revision and state throughout the manuscript that DPPN is trained for a 60-minute horizon, while the 70–120-minute results evaluate its recursive extension beyond that horizon. We agree that a sensitivity study over observation-window length and frame interval would add insight. Each setting requires retraining the corresponding models, so we list observation-window length, frame interval, and direct training for longer horizons as future work.

Reviewer_SR6XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper introduces Discrete Probabilistic Precipitation Nowcasting (DPPN), a framework that formulates probabilistic precipitation nowcasting as conditional generation over discrete radar tokens. Specifically, the method employs a pretrained VQ-VAE to encode continuous radar observations into discrete representations and adopts a discrete diffusion model to generate future precipitation tokens. The framework supports two forecasting formulations: direct generation conditioned on historical observations and cascaded generation conditioned on deterministic forecasts.

The authors further incorporate radar-space supervision and propose a data-informed corruption path to control the transitions between discrete token states during the diffusion process. Experiments on the SEVIR dataset demonstrate competitive forecasting performance compared with several continuous diffusion and flow-matching baselines. The proposed approach also achieves a smaller model size and reports lower computational costs under the evaluated implementations.

Overall, the paper presents an interesting application of discrete generative modeling to precipitation nowcasting. The experimental results show some promising advantages, particularly in pooled-scale precipitation event detection and computational efficiency. However, I have several concerns regarding the methodological novelty, experimental comprehensiveness, and the validity of some of the paper's central claims.

Critical strengths
  1. Exploration of discrete diffusion for precipitation forecasting. The paper investigates an alternative to continuous diffusion by representing radar observations as discrete tokens and modeling their conditional evolution. This provides an interesting perspective on probabilistic precipitation forecasting.

  2. Flexible forecasting formulations. The framework supports both direct generation from historical observations and cascaded generation conditioned on deterministic predictions, offering flexibility in how probabilistic forecasts are constructed.

  3. Promising computational efficiency. DPPN uses fewer trainable parameters than the evaluated continuous diffusion and flow-matching baselines. The reported results also suggest potential reductions in training and inference costs, although more controlled comparisons are needed to establish the extent of these advantages.

Critical weaknesses
  1. Limited methodological novelty and insufficient positioning against related work. The proposed framework appears to rely substantially on existing components, including VQ-VAE-based discretization, Transformer-based token modeling, and the discrete diffusion formulation adopted from previous work. Although the authors introduce task-specific radar-space supervision and a data-informed corruption path, the overall framework seems primarily to combine and adapt existing techniques to precipitation forecasting, rather than introduce a fundamentally new generative modeling mechanism. The authors should clarify the precise novelty of DPPN and explain how it differs from these related approaches. In particular, the broad claim of being the first discrete diffusion framework for probabilistic precipitation forecasting requires more careful qualification.
  2. Insufficient experimental coverage and generalization analysis. The experimental evaluation is restricted to a single dataset, SEVIR, with a fixed spatial resolution and forecasting horizon. Although the dataset provides a reasonable number of samples, the current experiments do not sufficiently establish whether the proposed discrete representation and diffusion mechanism generalize across different radar datasets, geographical regions, precipitation regimes, or forecasting settings. Given the claimed novelty of applying discrete diffusion to precipitation forecasting, a more comprehensive evaluation would be expected. In particular, additional experiments on independent datasets and different forecasting configurations would help establish the robustness and broader applicability of the proposed framework.
  3. Insufficient ablation studies of the key methodological components. Although the paper provides several ablation experiments, the current analysis does not sufficiently isolate the contributions of its central components.
  4. Incomplete evaluation of probabilistic forecasting quality. The paper reports CSI, mCSI, CRPS, HSS, and SSIM. While these metrics provide useful information about event detection and reconstruction quality, they do not fully characterize the quality of a probabilistic forecasting model.
Questions for authors
  1. Necessity of discrete diffusion: Have the authors compared DPPN against autoregressive and masked-token generation using the same tokenizer and a comparable Transformer backbone? If not, what evidence establishes that diffusion is particularly beneficial for this task?
  2. Effectiveness of the prescribed diffusion path: The proposed path demonstrates earlier token-transition stabilization, but the final forecasting metrics are not consistently comparable to those of the uniform path. Can the authors demonstrate improved sampling efficiency at matched forecasting quality and computational budgets?
1 comment
AuthorsOpenReview ↗

Response to Reviewer SR6X

We sincerely thank the reviewer for their time and feedback. Upon careful reading, we noticed that a few comments appear to refer to methods, datasets, or experiments that are not part of our manuscript, and we suspect they may have been inadvertently associated with our submission. To avoid misinterpreting the reviewer's intent, we have not addressed these specific points individually.

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

This paper proposes DPPN, a conditional discrete diffusion framework for probabilistic precipitation nowcasting. It tokenizes radar fields using a frozen vector-quantized representation and supports both direct forecasting from historical observations and cascaded forecasting conditioned on an Earthformer prediction. The method also introduces radar-space supervision and explores a data-informed nonuniform diffusion path derived from forecast–target token statistics. Experiments on SEVIR show competitive forecasting performance, particularly for pooled CSI and heavy precipitation at longer lead times, with a relatively compact trainable denoiser. The prescribed diffusion path leads to earlier token stabilization, although its practical quality–efficiency benefit is not yet fully established.

Critical strengths
  1. DPPN provides a unified discrete diffusion framework for precipitation nowcasting with two conditioning strategies. Radar-space supervision improves native-resolution mCSI.

  2. DPPN achieves strong pooled CSI and heavy-precipitation forecasting results, particularly at longer lead times. At 60 minutes, its CSI-219 outperforms CasCast and FlowCast.

Critical weaknesses
  1. DPPN and comparable methods use different ensemble fusion strategies. The feature-layout ablation also changes the fusion strategy, making it difficult to attribute performance improvements to the proposed model.

  2. Table 4 compares different sampling steps using inconsistent evaluation settings. Moreover, the prescribed path substantially reduces Pool16 mCSI, raising concerns about the claimed efficiency improvement without sacrificing forecasting quality.

Questions for authors
  1. Could the authors clarify the ensemble fusion strategies used in the main comparisons and ablation studies? Would DPPN's advantages in pooled and heavy-precipitation CSI remain under identical aggregation settings?

  2. Are the 30/48/64-step results in Table 4 complete sampling runs or intermediate states of a 100-step trajectory? Could the authors provide quality and inference-time comparisons under consistent settings, including matched additional training budgets for the uniform and prescribed paths?

  3. Appendix A.1 evaluates reconstruction using both top- and bottom-level representations, while DPPN models only top-level tokens. How are bottom-level representations obtained during inference? Could the authors report reconstruction quality using the actual DPPN decoding pipeline?

  4. Under what conditions is exact path following guaranteed, given the finite terminal sharpness in Eq. (20) and the finite-step approximation in Eq. (25)? Could the authors clarify the distinction between the theoretical guarantee and practical sampling?

6 comments
AuthorsOpenReview ↗

Overall response to reviewer fPpW

Thank you for your constructive comments and careful review of our paper. We clarify our ensemble aggregation approach and re-evaluate the baselines under the same aggregation rule. We also extend the experiments to the MeteoNet dataset. Across aggregation methods and datasets, our discrete diffusion model consistently yields better forecasts at coarser spatial scales. We clarify how our hierarchical VQ-VAE differs from original VQ-VAE-2: DPPN generates only top-level tokens, and the decoder infers the bottom-level codes from them. We further investigated the sampling protocols and the conditions for exact path following under finite sharpness and finite-step approximation. Our results show that the finite-step error of the prescribed path is small. The practical sampler stays close to the prescribed path throughout sampling step, with a small total variation (TV).

AuthorsOpenReview ↗

Response to reviewer comment Q1

Ensemble aggregation in the main comparison

CasCast, FlowCast, and FREUD+LSM-L average their decoded ensemble members pixel-wise. This is simple and robust but smooths localized intense echoes. DPPN's discrete codes cannot be averaged, so we also aggregate decoded fields, using a simple percentile-based pixel-wise fusion that applies to any generative model.

DPPN-cas and DPPN-direct use the P50/P60 fusion of Section 4.1: at each pixel, we take the 60th percentile (P60) across ensemble members if it is at least 74 on the VIL scale, and the median (P50) otherwise. We re-evaluated the three main baselines on SEVIR under this rule.

ModelParams (M)mCSI Pool1mCSI Pool4mCSI Pool16CSI-181 Pool16CSI-219 Pool16
CasCast [ICML 24]309.40.41270.46590.56760.44170.3166
FREUD+LSM-L [CVPR 26]473.00.41680.47790.56580.43070.3375‾\underline{0.3375}
FlowCast [ICLR 26]230.40.44150.4863‾\underline{0.4863}0.57190.44860.3036
DPPN-direct92.60.41140.46830.6095‾\underline{0.6095}0.4905‾\underline{0.4905}0.3194
DPPN-cas1020.4186‾\underline{0.4186}0.50490.63480.52760.4182

DPPN-cas leads the strongest baseline in mCSI by 0.0186 at Pool4 and 0.0629 at Pool16, while FlowCast leads at native resolution. The fusion also improves pooled strong-echo CSI: DPPN-cas reaches Pool16 CSI of 0.5276 at VIL 181 and 0.4182 at VIL 219, exceeding the strongest baseline by 0.0790 and 0.0807, respectively.

SEVIR with pixel-wise mean aggregation for all models. We also compare all models using the pixel-wise ensemble mean.

ModelParams (M)mCSI Pool1mCSI Pool4mCSI Pool16CSI-181 Pool16CSI-219 Pool16
CasCast [ICML 24]309.40.4253‾\underline{0.4253}0.48920.59540.47700.3541
FREUD+LSM-L [CVPR 26]473.00.41510.46560.54290.37900.2593
FlowCast [ICLR 26]230.40.44100.48130.56180.42600.2795
DPPN-cas1020.42210.4847‾\underline{0.4847}0.5774‾\underline{0.5774}0.4580‾\underline{0.4580}0.3444‾\underline{0.3444}

With mean aggregation, CasCast has the highest pooled mCSI, and DPPN-cas ranks second in Pool4/Pool16 mCSI and at both Pool16 strong-echo thresholds. DPPN-cas obtains mHSS of 0.5489, 0.6105, and 0.6717 at Pool1, Pool4, and Pool16, respectively.

MeteoNet. As requested, we also evaluate on MeteoNet with two fusions of eight ensemble members: the pixel-wise mean, and our percentile fusion (P60 where it is at least 35, P50 otherwise). With fusion, DPPN reaches Pool16 CSI-M of 0.4678 and CSI-47 of 0.3114, versus 0.4420 and 0.2357 for CasCast with the same EarthFormer deterministic estimator, showing that DPPN remains competitive on pooled strong-echo detection.

ModelCSI-M Pool1CSI-M Pool4CSI-M Pool16CSI-47 Pool1CSI-47 Pool4CSI-47 Pool16CRPS ↓
ConvLSTM [NeurIPS 15]0.30080.30500.34650.09820.10910.15880.0218
PredRNN [NeurIPS 17]0.29140.30030.34020.08230.09900.14620.0214
PhyDNet [CVPR 20]0.31200.31240.33560.11060.11570.14820.0216
SimVP [CVPR 22]0.30170.31430.35770.09970.11340.15990.0218
LDM [CVPR 22]0.21310.21910.23690.03590.04070.05520.0209
EarthFormer [NeurIPS 22]0.28310.28550.31540.07870.08720.12080.0224
PreDiff [NeurIPS 23]0.25460.26680.29350.04900.05940.08670.0197
NowcastNet [Nature 23]0.29550.32320.37340.12360.15210.21150.0277
CasCast (EarthFormer) [ICML 24]0.31560.36500.44200.12040.15630.23570.0180
exPreCast [ICLR 26]†0.28610.34520.4446‾\underline{0.4446}0.08560.13930.2525--
DPPN (Earthformer)-mean0.3323‾\underline{0.3323}0.3834‾\underline{0.3834}0.41760.1290‾\underline{0.1290}0.19400.2657‾\underline{0.2657}0.0165
DPPN (Earthformer)-fusion0.33690.38850.46780.14140.1937‾\underline{0.1937}0.31140.0165

† exPreCast scores are from the Table 3 of Song et al., Extreme Weather Nowcasting via Local Precipitation Pattern Prediction (ICLR 2026), which uses 416×416416\times416 crops and overlapping max pooling with stride k/4k/4 for kernel size kk. FREUD+LSM-L [CVPR 2026] reports mCSI of 0.1417 and CRPS of 0.0193 on MeteoNet under its own split; as its code and checkpoints are unavailable, we cannot run a matched comparison.

AuthorsOpenReview ↗

Response to comments from Reviewer fPpW

Sampling budgets, inference time, and matched training

What Table 4 reports. For each path, the 30-, 48-, and 64-step columns are checkpoints from a single reverse trajectory with a fixed total of 100 steps. At each checkpoint, we passed the current token IDs of one ensemble member straight to the decoder, without applying greedy projection. We used these checkpoints to examine how the trajectory evolves. The submitted caption called them “different sampling budgets,” and the 100-step column reported an eight-member forecast fused after terminal greedy projection. We will revise the caption and report the directly decoded single-member endpoint alongside the intermediate states.

When the sampled tokens stabilize. In the 100-step runs, the prescribed trajectory completed 95% of its observed token transitions by step 32, versus step 64 for the uniform trajectory (see the token jump-rate figure). After step 64, the prescribed trajectory made few further updates, whereas the uniform trajectory kept revising tokens until the final step.

Forecast quality and inference time in complete runs. We ran both samplers for the full 100 steps and fused each eight-member ensemble with the same rule. Sampling times are per member and include terminal greedy denoising.

SamplerPool1 mCSI (eight-member fusion)Single-member sampling time (ms/sample)Time saved vs. Uniform (ms/sample)
Prescribed0.4144245.785.3 (25.8%)
Uniform0.4196331.0— (baseline)

With the same 100-step budget, the prescribed sampler cuts single-member sampling time from 331.0 to 245.7 ms/sample (a 25.8% reduction), while its fused Pool1 mCSI is only 0.0052 lower than uniform (0.4144 vs. 0.4196). Together with its earlier token stabilization, this shows that the prescribed trajectory offers a favorable accuracy–efficiency trade-off: a marginal loss in native-resolution skill for a substantial reduction in inference cost.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

We thank the reviewers and the AC for their time and feedback. We have uploaded a revised manuscript. The main changes are summarized below, with details in the individual responses.

  1. Ensemble aggregation under identical settings. We re-evaluated CasCast, FREUD+LSM-L, and FlowCast on SEVIR with the same P50/P60 percentile fusion used by DPPN. DPPN-cas (102M parameters) exceeds the strongest baseline in mCSI by 0.0186 at Pool4 and 0.0629 at Pool16, and in Pool16 CSI-181/CSI-219 by 0.0790/0.0807, while FlowCast leads at native resolution. With pixel-wise mean aggregation for all models, DPPN-cas ranks second only to CasCast (309M) on pooled mCSI and pooled strong-echo CSI.

  2. Additional dataset. We trained the tokenizer and DPPN-cas on MeteoNet (Météo-France radar; 12 → 12 frames at 5-min intervals). Using the same EarthFormer deterministic forecast, DPPN reaches Pool16 CSI-M of 0.4678 and CSI-47 of 0.3114, versus 0.4420 and 0.2357 for CasCast, and obtains the lowest CRPS (0.0165 vs. 0.0180 for CasCast).

  3. Complete and transparent metric reporting. Table 1 is now split into (a) published SEVIR benchmark results (CRPS and mCSI at Pool1/4/16) and (b) a complete evaluation (SSIM, CRPS, HSS, mCSI, and 60-min mCSI) of all models with available checkpoints, so "–" no longer stands for metrics that cannot be evaluated.

  4. Forecast horizon. We now state explicitly that the main setting forecasts 60 min (7 → 6 frames at 10-min intervals). Extending the trained model autoregressively to 120 min without retraining, DPPN-cas outperforms EarthFormer at every lead time and pooling scale, and the Pool16 gap grows from 0.0255 at 70 min to 0.0711 at 120 min. Sensitivity to the observation window, frame interval, and direct long-horizon training is listed as future work.

  5. Prescribed path: sampling protocol and efficiency. The 30/48/64-step columns in Table 4 are checkpoints of a single 100-step trajectory, and we have corrected the caption. The prescribed sampler completes 95% of its token transitions by step 32, versus step 64 for the uniform sampler. In complete 100-step runs with identical fusion, it reduces per-member sampling time by 25.8% (245.7 vs. 331.0 ms) at a cost of 0.0052 in Pool1 mCSI.

  6. Tokenizer decoding. DPPN generates only top-level tokens, and our restructured decoder infers the bottom-level codes from them. All reconstructions, including Appendix A.1, use this top-only pipeline. Compared with a flat 16×16 VQ-VAE using the same 256 tokens per frame, it improves mCSI from 0.5694 to 0.6032 and CSI@219 from 0.2481 to 0.3428.

  7. Theoretical guarantee vs. practical sampling. The revised appendix separates three effects: (i) the continuous-ρ process follows the prescribed path exactly when the exact path posterior is used; (ii) the finite-β_max endpoint error is bounded in TV by ε_term = 1.63×10⁻⁵ at β_max = 20; and (iii) the finite-step error halves as the number of steps doubles (TV of 0.0072, 0.0036, and 0.0018 at 100, 200, and 400 steps). With the learned denoiser, the practical sampler stays within a TV of 0.05 of the prescribed marginals at all measured checkpoints.

We hope these revisions address the concerns raised, and we welcome any further questions during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.