acceptodds
Withdrawn from ICLR 2027

NowcastDiT: Diffusion Transformers are Effective Precipitation Nowcasters

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Precipitation nowcasting demands accurate short-term forecasts under strong spatiotemporal variability. Diffusion models are well suited to modeling complex precipitation distributions, yet existing approaches often introduce increasingly specialized designs, leaving the capability of a standard diffusion architecture underexplored. We show that a standard Diffusion Transformer already provides a simple and scalable foundation for precipitation nowcasting, with domain-specific requirements accommodated naturally within its design space. Based on this principle, we develop NowcastDiT and instantiate this flexibility through two complementary adaptations: a dynamics-aware noise prior for temporally coherent forecasts, and end-to-end reinforcement learning with stage-aware rewards for meteorological skill. Experiments on SEVIR and MRMS benchmarks show that NowcastDiT achieves state-of-the-art performance in both perceptual quality and meteorological skill. These results suggest that standard DiT can serve as a effective foundation for precipitation nowcasting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Precipitation Nowcasting, Diffusion Models, AI for Science, Physical Ingellegence

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5zhCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose NowcastDiT, a latent Diffusion Transformer model for precipitation nowcasting. The model is largely based on a standard DiT, with the goal of using a general architecture rather than one specifically designed for precipitation nowcasting. The two components added to the standard DiT are DyPro, which aims to improve temporal coherence through a prior with temporally correlated noise, and timestep-aware GRPO post-training, which shifts the reward emphasis from coarse structure to fine details along the denoising trajectory. Results on SEVIR and MRMS show strong performance compared with the included deterministic and generative baselines across skill and perceptual metrics.

Critical strengths
  1. Clear and well-motivated paper. The paper is well written and easy to follow. Having a general architecture work well for precipitation nowcasting, rather than relying on a heavily specialized one, is a strong motivation. This direction has been successful in the past (i.e. ConvLSTM, Earthformer, SimVP…).
  2. Strong results and informative ablations. NowcastDiT achieves strong results across skill and perceptual metrics. Table 6 shows that the standard DiT is already a strong baseline, that 3D RoPE and QK-Norm improve it further, and that both proposed components (DyPro and RL) provide additional gains.
Critical weaknesses
  1. Novelty and related work. I find the novelty and the standard DiT argument somewhat weak. Applying a standard DiT to precipitation nowcasting is useful but is mainly an application of an existing architecture. The use of general architectures for precipitation nowcasting is also well established (e.g., ConvLSTM, Earthformer, SimVP). Similarly, latent diffusion has already been explored (PreDiff, LDCast, FREUD), and flow matching has also been explored (FlowCast - ICLR 2026). Importantly, several of these closely related works, including FlowCast, LDCast and FREUD, are not discussed. Finally, the best performing NowcastDiT is not just a standard DiT, but relies on DyPro and precipitation-specific RL rewards. I therefore think that the claims around architectural simplicity should be reduced and the paper should focus more on its actual contributions.
  2. Contributions and experiments. Beyond applying DiT to precipitation nowcasting, the main contributions are DyPro for temporal coherence and RL post-training for improving forecast quality. The ablations show that both components help, but they are only compared against the authors’ own baseline, rather than alternative approaches for temporal coherence or post-training discussed in the related work. The evaluation also misses important recent methods, including FlowCast, LDCast, PercpCast, FREUD, and DuoCast. Finally, the evaluation of a probabilistic model would benefit from metrics such as CRPS, calibration, and ensemble diversity.
Questions for authors
  1. How does NowcastDiT compare with recent and closely related methods such as FlowCast, LDCast, PercpCast, FREUD and DuoCast?
  2. How do DyPro and RL compare with alternative approaches for temporal coherence and post-training?
  3. Can the authors report probabilistic metrics such as CRPS, calibration, and ensemble diversity?
  4. How does NowcastDiT compare with the baselines in terms of model size and training / inference costs, given the scaling results in Figure 5a?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 5zhC (Part I)

We sincerely thank Reviewer 5zhC for the insightful review and constructive comments. Below, we address your concerns in detail.

[W1.1, W1.2] Contribution of the standard DiT foundation

We acknowledge that applying a standard DiT [1] is architecturally straightforward. Yet we believe our contribution also lies in systematically validating its effectiveness as a general-purpose foundation for precipitation nowcasting.

  • Establishing a standard architecture is itself a contribution for subsequent research. The development of ViT illustrates how exploring a standard Transformer alongside specialized hybrid designs can unlock architectural potential and enable subsequent advances in scaling, implementation efficiency, and hardware optimization. NowcastDiT explores this standard-architecture paradigm for precipitation nowcasting.
  • How far a standard Transformer can go in precipitation nowcasting remains insufficiently explored. While ConvLSTM, Earthformer [2], and SimVP have demonstrated the use of “general architectures” for precipitation nowcasting, the possibility of a standard Transformer serving as a general-purpose foundation remains underexplored, especially in generative nowcasting. NowcastDiT not only establishes the feasibility of a standard Transformer foundation but also thoroughly evaluates its capabilities through extensive baseline comparisons, architectural ablations, and model scaling.
    • Discussion on FREUD. While the concurrent work FREUD also adopts a standard DiT backbone, it focuses on specialized tokenizer design and evaluates the backbone only with its proposed tokenizer, leaving its performance with a standard VAE unexplored—a gap that NowcastDiT directly addresses. Furthermore, FREUD shows mixed results against existing baselines, while NowcastDiT outperforms it on all four reported metrics in our additional comparison (see our response to [W2.2, Q1]).
  • The standard backbone delivers strong results independently of DyPro and RL. The standard DiT achieves competitive performance on both datasets. Incorporating general-purpose DiT techniques such as 3D RoPE and QK-Norm further improves performance, as shown in Table 6 in appendix, outperforming all external baselines in our main comparison before adding DyPro or RL.

Taken together, these findings establish a standard DiT as an effective foundation for precipitation nowcasting, enabling subsequent research to build on advances in general-purpose generative modeling.

[W1.3, W2.2, Q1] Additional Baseline Comparisons

We acknowledge the omission of these baselines from the main comparison. LDCast, PercpCast, FlowCast, and DuoCast follow the evaluation protocol established by AlphaPre [3], whose test split differs from the DiffCast [4] protocol used in our main experiments. For a fair comparison, we re-evaluate NowcastDiT under the same protocol.

MethodCSI-M ↑CSI-181 ↑CSI-219 ↑HSS ↑SSIM ↑LPIPS ↓
AlphaPre [3]0.32590.13320.05450.41100.68840.2855
LDCast [5]0.32760.14230.06850.41870.69510.1879
PercpCast [6]0.32010.13560.05920.40260.66490.1960
FlowCast [7]0.33470.17260.08730.43370.69570.1883
DuoCast [8]0.33750.18180.10740.43180.68270.2263
NowcastDiT0.36910.21320.12200.47690.70370.1629

FREUD [9] uses a different temporal configuration, predicting 12 frames from 13 input frames, together with a higher spatial resolution of 384×384. We therefore separately retrain NowcastDiT under the same 13-to-12 configuration and evaluate both methods at 128×128 for a fair comparison.

MethodCRPS ↓SSIM ↑HSS ↑CSI-M ↑
FREUD0.01830.73310.48470.3800
NowcastDiT0.01720.75480.52110.4083

Across both evaluation settings, NowcastDiT outperforms all compared methods on reported metric, demonstrating consistent advantages in meteorological skill, perceptual quality, and probabilistic forecast quality.

[W1.4] Reducing overclaims of “simplicity”

We acknowledge the reviewer’s concern about the overemphasis on architectural simplicity. We thank the reviewer for this suggestion and have removed repetitive references to “standard” and “simple,” retaining them only where appropriate.

AuthorsOpenReview ↗

Response to Reviewer 5zhC (Part II)

[W2.1.1, Q2.1] Alternative temporal-coherence methods

To supplement comparison with fixed-correlation noise, we conduct experiments on schedules of Progressive and Mixed noise from PYoCo [10], and Dynamical Diffusion [11]. Our DyPro achieves the highest CSI, HSS, and SSIM scores and the second-lowest LPIPS, demonstrating strong overall forecasting performance.

Noise constructionCSI ↑CSI-181 ↑CSI-219 ↑HSS ↑LPIPS ↓SSIM ↑
IID Gaussian0.30120.10820.06760.36470.13170.6812
Progressive Noise0.30410.10890.06120.39140.15390.6930
Mixed Noise0.31220.11730.06790.40090.15120.7052
Dynamical Diffusion0.31480.11470.06900.40330.15030.7046
DyPro (ours)0.31950.12040.07520.40920.14860.7099

[W2.1.2, Q2.2] Alternative post-training methods

We acknowledge the concerns on alternative post-training methods. Notably, for post-training, established alternatives for diffusion-based precipitation nowcasting remain limited. To broaden this comparison, we consider general post-training methods [12-14] in diffusion models. As shown in the table below, post-training of NowcastDiT achieves the best CSI, HSS, SSIM and the second-lowest LPIPS, demonstrating strong overall forecasting performance.

Post-training methodCSI ↑HSS ↑SSIM ↑LPIPS ↓
w/o post-training0.31950.40920.70990.1486
FlowGRPO0.32020.41200.71020.1505
DanceGRPO0.32120.41160.71410.1495
MixGRPO0.32180.41250.71460.1473
NowcastDiT0.32540.41600.71940.1479

[Q3] Scope of probabilistic evaluation

We acknowledge the omission of ensemble metrics. Our evaluation protocol directly follows that established in [4], which focuses on deterministic forecast quality and does not include these metrics. To extend this evaluation, we report ensemble CRPS, ensemble-mean RMSE, and spread for all three configurations on SEVIR. These metrics are computed using four ensemble members, with predictions and observations normalized to [0,1]. Following [9], we disable classifier-free guidance (CFG) when evaluating probabilistic forecast quality.

ConfigurationCSI ↑CRPS ↓RMSE ↓Spread
DiT0.27060.023720.067230.05362
DiT + DyPro0.29340.021540.061140.04114
DiT + DyPro + RL0.31090.022140.062560.03855

To examine calibration, we compare rank-histogram frequencies for all three configurations. The uniform reference is 20% per bin.

ConfigurationRank 0Rank 1Rank 2Rank 3Rank 4
DiT34.55%18.49%14.42%14.04%18.50%
DiT + DyPro31.57%18.98%14.72%14.38%20.36%
DiT + DyPro + RL40.16%15.46%11.84%12.16%20.38%

The results highlight two findings:

  • DyPro improves CSI, CRPS, and RMSE and produces a more balanced rank distribution, showing that its noise-prior design benefits multiple aspects of forecast quality.
  • RL substantially improves individual-forecast skill with a trade-off in ensemble metrics. This is consistent with its focus on per-forecast meteorological and perceptual rewards. These results suggest using DiT with DyPro when prioritizing ensemble forecast quality, and the post-trained model when prioritizing individual-forecast skill.
AuthorsOpenReview ↗

Response to Reviewer 5zhC (Part III)

[Q4] Model size and computational costs

To contextualize the scaling results in Figure 5a, we report model sizes, training costs, and inference throughput below. Detailed measurement protocols and configurations are provided in the appendix.

MethodParams (M)Training cost (GPU-h) ↓Throughput (seq/s) ↑NFE/sequence ↓CSI ↑
SimVP11.06260.3942.80—0.2951
AlphaPre89.03467.1320.82—0.2980
DiffCast49.36212.520.3910000.2977
CasCast81.71 + 349.36525.162.99200.2905
NowcastDiT-S81.71 + 32.72150.905.50200.3119
NowcastDiT-B81.71 + 130.32228.733.41200.3195

NowcastDiT-S outperforms all listed baselines in CSI while requiring the lowest measured training cost, and achieves higher inference throughput than DiffCast and CasCast. Scaling to NowcastDiT-B further improves CSI at increased computational cost. Deterministic models offer higher inference throughput, while NowcastDiT delivers stronger forecast skill.

References

[1] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. ICCV, 2023.

[2] Zhihan Gao et al. Earthformer: Exploring Space-Time Transformers for Earth System Forecasting. NeurIPS, 2022.

[3] Kenghong Lin et al. AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting. CVPR, 2025.

[4] Demin Yu et al. DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting. CVPR, 2024.

[5] Jussi Leinonen et al. Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification. arXiv:2304.12891, 2023.

[6] Wenzhi Feng et al. Perceptually Constrained Precipitation Nowcasting Model. ICML, 2025.

[7] Bernardo Perrone Ribeiro and Jana Faganeli Pucer. FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching. ICLR, 2026.

[8] Penghui Wen et al. DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting. AAAI, 2026.

[9] Johannes Schusterbauer et al. Probabilistic Precipitation Nowcasting with Rectified Flow Transformers. CVPR, 2026.

[10] Songwei Ge et al. Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models. ICCV, 2023.

[11] Xingzhuo Guo et al. Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models. ICLR, 2025.

[12] Jie Liu et al. Flow-GRPO: Training Flow Matching Models via Online RL. NeurIPS, 2025.

[13] Zeyue Xue et al. DanceGRPO: Unleashing GRPO on Visual Generation. arXiv:2505.07818, 2025.

[14] Junzhe Li et al. MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE. ECCV, 2026.

AuthorsOpenReview ↗

Request of Reviewer's Attention and Feedback

Dear Reviewer,

This is a kind reminder that three days remain in the discussion period. We would appreciate your feedback on whether our response has addressed your concerns.

Following your suggestions, we have:

  • Clarified the contribution of exploring standard DiT foundation for nowcasting .
  • Added comparisons with LDCast, PercpCast, FlowCast, DuoCast, and FREUD.
  • Reduced repetitive claims about architectural simplicity.
  • Compared alternative noise priors and post-training methods.
  • Added ensemble metrics and calibration diagnostics.
  • Reported model sizes and training/inference costs.

Thanks again for your valuable review. We look forward to your feedback and are happy to clarify any remaining points.

Reviewer_589pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes NowcastDiT, a precipitation nowcasting framework that keeps a standard Diffusion Transformer backbone and introduces two lightweight, task-aligned adaptations: (i) a dynamics-aware noise prior (DyPro) that injects timestep-scheduled temporal correlation across forecast frames to promote temporal coherence, and (ii) an end-to-end reinforcement learning (GRPO-style) post-training scheme with timestep-aware rewards that shift emphasis from large-scale structure to fine-scale detail along the denoising trajectory. On SEVIR and MRMS benchmarks, the method reports state-of-the-art CSI/HSS (meteorological skill) and strong perceptual metrics (SSIM/LPIPS), with thorough ablations on model scaling, DyPro strength, CFG+RL interactions, and reward scheduling.

Critical strengths
  1. Simple and well-formulated framework. The paper presents a straightforward DiT-based framework with two task-oriented adaptations. DyPro is mathematically well specified, including its covariance properties and corresponding flow-matching sampling rule. The timestep-aware GRPO reward also provides a reasonable approach to incorporating meteorological objectives into diffusion post-training.

  2. Competitive experimental results. The method demonstrates encouraging improvements in meteorological skill metrics on SEVIR and MRMS. The inclusion of a standard DiT baseline and ablations on noise correlation strength, CFG–RL interactions, and reward scheduling helps evaluate the effectiveness of the proposed components.

Critical weaknesses
  1. Limited methodological novelty. The proposed DyPro closely resembles existing temporally correlated noise priors. PYoCo (ICCV 2023) introduced a similar autoregressive noise construction, while Dynamical Diffusion (ICLR 2025) has already explored timestep-dependent temporal correlations. The main contribution therefore appears to be the adaptation of existing mechanisms to flow matching rather than a fundamentally new noise prior. Similarly, the timestep-aware reward relies on a heuristic linear interpolation between predefined objectives, offering relatively limited methodological novelty.

  2. Insufficient ablations to establish the key contributions. Although the paper compares DyPro against independent Gaussian noise and investigates different correlation strengths, it lacks comparisons with fixed-correlation priors and existing timestep-dependent correlation schedules. Consequently, it remains unclear whether the improvements come from temporal noise correlation itself or the proposed scheduling mechanism. Moreover, while the timestep-aware reward outperforms static weighting, the assumed distinction between early-stage structural generation and late-stage detail refinement is not directly validated.

  3. Incomplete experimental validation. The paper emphasizes temporal coherence but does not provide direct quantitative evaluations of temporal consistency. Probabilistic forecasting metrics, such as CRPS and calibration measures, are also missing despite the generative nature of the model. Furthermore, relevant recent methods, including DuoCast (AAAI 2026), are not included in the main comparisons, making the state-of-the-art claims less conclusive.

Questions for authors
  1. Novelty of DyPro. How does DyPro fundamentally differ from existing temporally correlated noise mechanisms, particularly PYoCo and Dynamical Diffusion? Could the authors provide controlled comparisons with fixed-correlation noise priors and existing timestep-dependent schedules to demonstrate the specific contribution of the proposed design?

  2. Effectiveness of temporal correlation. Since improving temporal coherence is a central motivation of DyPro, could the authors provide direct quantitative evaluations of temporal consistency? A per-lead-time comparison isolating DyPro's contribution would also help clarify where the proposed noise mechanism is most beneficial.

  3. Experimental completeness. Could the authors include comparisons with relevant recent methods, such as DuoCast (AAAI 2026), under consistent evaluation protocols, or clarify the differences that prevent direct comparison? Additionally, given the probabilistic nature of the model, could the authors report uncertainty calibration metrics such as CRPS to assess the effect of RL post-training?

I am open to raising my score if the authors can provide convincing evidence that DyPro offers meaningful advantages over existing temporally correlated noise priors, particularly through controlled comparisons with fixed-correlation baselines and prior timestep-dependent schedules. A clearer justification of the methodological novelty and additional evidence supporting the claimed temporal coherence improvements would also help address my main concerns.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 589p (Part I)

We sincerely thank Reviewer 589p for the insightful review and constructive comments. We appreciate your recognition of our framework's formulation and forecasting performance. Below, we address your concerns in detail.

[W1.1, W2.1, Q1] Clarifying novelty of noise prior

We acknowledge reviewer's concern about the novelty of noise prior. We would like to clarify that DyPro is not an existing mechanism compared with existing works, and not an intentional adaptation from diffusion to flow models.

  • Differences from PYoCo [1]: PYoCo uses a single factor, keeping the noise correlation constantly fixed across all denoising steps. In contrast, DyPro adopts a different motivation that preserves correlation during early denoising but allows frame-specific variations during later denoising, leading to a timestep-dependent correlation different from PYoCo.
  • Differences from Dynamical Diffusion [2]: Dynamical Diffusion, while also introducing a timestep-aware schedule, mixes both clean signals and noise. DyPro instead mixes only the noise while keeping the clean target and flow-matching objective unchanged, simplifying training and resulting in a different mathematical formulation.
  • Not an adaptation of a diffusion schedule to flow models: Diffusion models and flow models are in fact unified under the same formulation by [3]. While our formulation is derived within the flow-model framework, we do not view this as an adaptation to flow models; we adopt this formulation only for its widely known efficiency.

To supplement comparison, we conduct experiments on the corresponding baselines. Our DyPro achieves the highest CSI, HSS, and SSIM scores and the second-lowest LPIPS, demonstrating strong overall forecasting performance.

Noise constructionCSI ↑CSI-181 ↑CSI-219 ↑HSS ↑LPIPS ↓SSIM ↑
IID Gaussian0.30120.10820.06760.36470.13170.6812
Progressive Noise0.30410.10890.06120.39140.15390.6930
Mixed Noise0.31220.11730.06790.40090.15120.7052
Dynamical Diffusion0.31480.11470.06900.40330.15030.7046
DyPro (ours)0.31950.12040.07520.40920.14860.7099

[W1.2] Clarifying novelty of post-training

We acknowledge the concerns on the predefined, prior form of the timestep-aware reward. Yet, we would like to clarify the novelty of our post-training:

  • Novelty on aligning meteorological metrics to diffusion timesteps: Notably, our “linear interpolation” is a weighting scheme informed by how different denoising stages affect different metrics, rather than constant balancing. To the best of our knowledge, no prior meteorological work has identified how metrics with different properties should be emphasized across diffusion timesteps. We adopt a linear schedule for simplicity, while more flexible weighting schedules could be explored in future work.
  • Novelty in instantiating timestep-aware rewards: To the best of our knowledge, NowcastDiT is the first to make the reward formulation timestep-aware across meteorological forecasting and diffusion modeling. This design both provides a simple approach to optimizing meteorological metrics and offers a methodological insight that reward emphasis can be aligned with the distinct roles of different denoising stages.
AuthorsOpenReview ↗

Response to Reviewer 589p (Part II)

[W2.2, W3.1, Q2] Temporal coherence and long lead-time evaluation

To supplement detailed analysis according to direct temporal consistency metrics, we assess temporal coherence using signed frame differences: Δxt=xt+1−xt\Delta x_t=x_{t+1}-x_t for references and Δx^t=x^t+1−x^t\Delta \hat{x}_t=\hat{x}_{t+1}-\hat{x}_t for predictions, and define the following two metrics:

  • Tendency Wasserstein Distance (TWD) measures their distributional agreement using the Wasserstein-1 distance:

    TWD=1T−1∑t=1T−1W1![P(Δx^t),P(Δxt)].\mathrm{TWD}=\frac{1}{T-1}\sum_{t=1}^{T-1}\mathcal{W}_1\\!\left[\mathcal{P}(\Delta\hat{x}_t),\mathcal{P}(\Delta x_t)\right].

  • Tendency Error (TE) measures their positional agreement via averaged ℓ1\ell_1 distance:

    TE=1T−1∑t=1T−1Eij[∣Δx^t,i,j−Δxt,i,j∣1].\mathrm{TE}=\frac{1}{T-1}\sum_{t=1}^{T-1}\mathbb{E}_{ij}\left[\left|\Delta\hat{x}_{t,i,j}-\Delta x_{t,i,j}\right|_1\right].

Detailed definitions and motivations are provided in the revised manuscript. We also include the alternative noise constructions discussed in [W1.1, W2.1, Q1].

Noise typeTWD ↓ (×10⁻³)TE ↓ (×10⁻³)
IID Gaussian5.00919.924
Progressive Noise3.809 (−23.96%)22.843 (+14.65%)
Mixed Noise3.654 (−27.05%)20.704 (+3.91%)
Dynamical Diffusion4.229 (−15.57%)20.280 (+1.79%)
DyPro (ours)3.633 (−27.47%)20.630 (+3.54%)

DyPro substantially improves the distributional match and achieves the lowest TWD, with a modest increase in pointwise error compared with IID Gaussian. This aligns with the motivation of DyPro: its early-denoising coupling promotes shared structure across frames, while later relaxation allows local growth and decay, a balance reflected in the closer match to the observed distribution of frame-to-frame changes.

Furthermore, we agree that a per-lead-time comparison is beneficial, as improved temporal coherence is expected to yield larger gains at longer lead times. We have added this comparison below. The results support this expectation: DyPro’s relative CSI gains over IID Gaussian noise become increasingly pronounced as lead time increases.

ModelCSI@20 min↑CSI@40 min↑CSI@60 min↑CSI@80 min↑CSI@100 min↑
IID Gaussian0.37650.28410.23970.21280.1923
AlphaPre (baseline SOTA)0.3972 (+5.50%)0.2837 (−0.14%)0.2351 (−1.92%)0.2059 (−3.24%)0.1838 (−4.42%)
DyPro (ours)0.4049 (+7.54%)0.3133 (+10.28%)0.2692 (+12.31%)0.2410 (+13.25%)0.2193 (+14.04%)

[W2.3] Timestep-aware rewards and the structure-to-detail motivation

We acknowledge the concerns regarding the explanation of our structure-to-detail design and clarify our motivation below:

  • Existing observations of diffusion models: Prior studies [4,5] have observed a progression from coarse content formation to fine-detail refinement during diffusion sampling. This motivates our gradual shift in reward emphasis from precipitation coverage and large-scale spatial patterns to local intensity variations and heavy-precipitation cores. We have included the corresponding discussion in our manuscript.
  • Case studies as direct evidence: To further examine this behavior in our setting, we provide decoded clean-prediction estimates at multiple denoising timesteps along the same sampling trajectory (Figure 13 on the last page). These show that precipitation coverage and large-scale spatial patterns emerge first, followed by increasingly refined local intensity variations and more pronounced heavy-precipitation cores.

[W3.2, Q3.1] Comparison with Recent Baselines

We acknowledge the omission of concurrent work DuoCast [6] from the main comparison. Notably, DuoCast (originating from AlphaPre [7]) uses a different test split from that of DiffCast [8] in our setup. For fair comparison, we re-evaluate NowcastDiT under their setup. The results are shown below. Compared with DuoCast's reported results, NowcastDiT consistently outperforms DuoCast.

MethodCSI-M ↑CSI-181 ↑CSI-219 ↑HSS ↑SSIM ↑
(AlphaPre)0.32590.13320.05450.41100.6884
DuoCast0.33750.18180.10740.43180.6827
NowcastDiT0.36910.21320.12200.47690.7037
AuthorsOpenReview ↗

Response to Reviewer 589p (Part III)

[W3.1, Q3.2] Ensemble Evaluation and Calibration

We acknowledge the omission of ensemble metrics. Our evaluation protocol directly follows that established in [8], which focuses on deterministic forecast quality and does not include these metrics. To extend this evaluation, we report ensemble CRPS, ensemble-mean RMSE, and spread for all three configurations on SEVIR. These metrics are computed using four ensemble members, with predictions and observations normalized to [0,1]. Following [9], we disable classifier-free guidance (CFG) when evaluating probabilistic forecast quality.

ConfigurationCSI ↑CRPS ↓RMSE ↓Spread
DiT0.27060.023720.067230.05362
DiT + DyPro0.29340.021540.061140.04114
DiT + DyPro + RL0.31090.022140.062560.03855

To examine calibration, we compare rank-histogram frequencies for all three configurations. The uniform reference is 20% per bin.

ConfigurationRank 0Rank 1Rank 2Rank 3Rank 4
DiT34.55%18.49%14.42%14.04%18.50%
DiT + DyPro31.57%18.98%14.72%14.38%20.36%
DiT + DyPro + RL40.16%15.46%11.84%12.16%20.38%

The results highlight two findings:

  • DyPro improves CSI, CRPS, and RMSE and produces a more balanced rank distribution, showing that its noise-prior design benefits multiple aspects of forecast quality.
  • RL substantially improves individual-forecast skill with a modest trade-off in ensemble metrics. This is consistent with its focus on per-forecast meteorological and perceptual rewards. These results suggest using DiT with DyPro when prioritizing ensemble forecast quality, and the post-trained model when prioritizing individual-forecast skill.

References

[1] Songwei Ge et al. Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models. ICCV, 2023.

[2] Xingzhuo Guo et al. Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models. ICLR, 2025.

[3] Michael Albergo et al. Stochastic Interpolants: A Unifying Framework for Flows and Diffusions. JMLR, 2025.

[4] Jonathan Ho et al. Denoising Diffusion Probabilistic Models. NeurIPS, 2020.

[5] Yogesh Balaji et al. eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers. arXiv:2211.01324, 2022.

[6] Penghui Wen et al. DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting. AAAI, 2026.

[7] Kenghong Lin et al. AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting. CVPR, 2025.

[8] Demin Yu et al. DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting. CVPR, 2024.

[9] Johannes Schusterbauer et al. Probabilistic Precipitation Nowcasting with Rectified Flow Transformers. CVPR, 2026.

AuthorsOpenReview ↗

Request of Reviewer's Attention and Feedback

Dear Reviewer,

This is a kind reminder that three days remain in the discussion period. We would appreciate your feedback on whether our response has addressed your concerns.

Following your suggestions, we have:

  • Clarified DyPro’s differences from existing noise priors and added corresponding comparisons.
  • Clarified the contribution of post-training that aligning meteorological rewards with denoising timesteps.
  • Added temporal metrics and per-lead-time CSI comparisons to evaluate DyPro.
  • Expanded the motivation for the reward schedule, supported by denoising visualizations.
  • Compared with DuoCast under the same evaluation setting.
  • Added ensemble metrics and calibration diagnostics.

Thanks again for your valuable review. We look forward to your feedback and are happy to clarify any remaining points.

Reviewer_b78COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents NowcastDiT, a precipitation nowcasting framework built on a standard DiT. It introduces a dynamics-aware noise prior to improve temporal coherence and GRPO-based post-training with timestep-aware meteorological and perceptual rewards. The method is evaluated on SEVIR and MRMS, predicting 20 future frames from 5 and 4 observations, respectively. Experiments report improvements in CSI and HSS over existing baselines, along with competitive perceptual quality. Additional ablations examine the effects of noise correlation, guidance, reinforcement learning, and reward design. The results suggest that a general-purpose DiT can effectively support precipitation nowcasting with targeted adaptations.

Critical strengths
  1. NowcastDiT achieves the highest aggregate CSI on both SEVIR and MRMS, with notable improvements at high precipitation thresholds.

  2. The method retains a standard DiT backbone, while ablation studies demonstrate the effectiveness of DyPro and RL post-training.

Critical weaknesses
  1. The static and timestep-aware rewards have different average weights under the window schedule, making it unclear whether the gains come from temporal adaptation or different reward emphasis.

  2. The paper lacks comparisons with fixed-correlation noise and direct temporal consistency metrics, making it difficult to verify the benefit of the proposed dynamic noise schedule.

Questions for authors
  1. Does timestep-aware weighting still outperform a static reward with matched average weights? A comparison with reversed or shuffled timestep assignments would help isolate the benefit of temporal adaptation.

  2. How does the proposed dynamic noise correlation compare with fixed-correlation noise or alternative schedules? Can the authors provide quantitative temporal consistency metrics to support the claimed benefits?

  3. Algorithm 4 averages Gaussian log probabilities across latent dimensions, whereas Equations 39–41 use the full joint log density. Which formulation is used in practice, and is any rescaling applied? Please also clarify the stochasticity schedule �0��0�.

  4. Table 6 shows that RL improves CSI from 0.3195 to 0.3240. Are these improvements consistent across multiple random seeds? Please provide variance or uncertainty estimates to demonstrate the robustness of the reported gains.

3 comments
AuthorsOpenReview ↗

Response to Reviewer b78C (Part I)

We sincerely thank Reviewer b78C for the insightful review and constructive comments. We appreciate your recognition of NowcastDiT's forecasting performance and the effectiveness of its components. Below, we address your concerns in detail.

[W1, Q1] Clarifying the benefit of temporal adaptation

We acknowledge the potential weight mismatch under the window schedule. To isolate the benefit of temporal adaptation, we discuss the following general reward formulation:

R=αtRnoise+βtRclean,R=\alpha_t R_{\mathrm{noise}}+\beta_t R_{\mathrm{clean}},

where NowcastDiT and the static schedule use (αt,βt)=(1−t,t)(\alpha_t,\beta_t)=(1-t,t) and (αt,βt)=(0.5,0.5)(\alpha_t,\beta_t)=(0.5, 0.5), respectively. We supplement with baselines

&\text{matched-static: } (\alpha_t,\beta_t) =(\mathbb{E}_{t\sim W}[1-t],\mathbb{E}_{t\sim W}[t])=(0.65,0.35), \\\\ &\text{reversed: } (\alpha_t,\beta_t) =(t,1-t), \\\\ &\text{shuffled: } (\alpha_{t_n},\beta_{t_n}) =\bigl(1-t_{\mathrm{perm}_N(n)},t_{\mathrm{perm}_N(n)}\bigr). \end{aligned}$$ where ${t_n}_{n=0}^{N-1}$ denotes the sequence of sampling timesteps and $\mathrm{perm}$ is a random permutation of indexes that remains fixed within each update interval. The table below reports the results under these schedules, with all other settings held constant: | Reward weighting | CSI ↑ | HSS ↑ | SSIM ↑ | LPIPS ↓ | | ------------------------ | ---------- | ---------- | ---------- | ---------- | | Static weighting | 0.3214 | 0.4120 | 0.7123 | 0.1481 | | Matched-static weighting | 0.3218 | 0.4125 | 0.7146 | **0.1473** | | Reversed weighting | 0.3230 | 0.4135 | 0.7162 | 0.1500 | | Shuffled weighting | 0.3222 | 0.4127 | 0.7148 | 0.1479 | | NowcastDiT (ours) | **0.3254** | **0.4160** | **0.7194** | 0.1479 | Our timestep-aware weighting achieves the highest CSI, HSS, and SSIM while attaining a joint second-best LPIPS score. These results support the benefit of adapting reward weights to sampling timesteps beyond matching their average values. ## [W2.1, Q2.1] Correlated-noise baselines To extend our comparison with correlated-noise baselines, we conduct experiments on schedules of Progressive and Mixed noise from PYoCo [1], and Dynamical Diffusion [2]. Our DyPro achieves the highest CSI, HSS, and SSIM scores and the second-lowest LPIPS, demonstrating strong overall forecasting performance. | Noise construction | CSI ↑ | CSI-181 ↑ | CSI-219 ↑ | HSS ↑ | LPIPS ↓ | SSIM ↑ | | --- | :---: | :---: | :---: | :---: | :---: | :---: | | IID Gaussian | 0.3012 | 0.1082 | 0.0676 | 0.3647 | **0.1317** | 0.6812 | | Progressive Noise | 0.3041 | 0.1089 | 0.0612 | 0.3914 | 0.1539 | 0.6930 | | Mixed Noise | 0.3122 | 0.1173 | 0.0679 | 0.4009 | 0.1512 | 0.7052 | | Dynamical Diffusion | 0.3148 | 0.1147 | 0.0690 | 0.4033 | 0.1503 | 0.7046 | | DyPro (ours) | **0.3195** | **0.1204** | **0.0752** | **0.4092** | 0.1486 | **0.7099** | ## [W2.2, Q2.2] Temporal coherence evaluation To supplement detailed analysis according to direct temporal consistency metrics, we assess temporal coherence using signed frame differences: $\Delta x_t=x_{t+1}-x_t$ for references and $\Delta \hat{x}_t=\hat{x}_{t+1}-\hat{x}_t$ for predictions, and define the following two metrics: - **Tendency Wasserstein Distance (TWD)** measures their distributional agreement using the Wasserstein-1 distance: $$\mathrm{TWD}=\frac{1}{T-1}\sum_{t=1}^{T-1}\mathcal{W}_1\\!\left[\mathcal{P}(\Delta\hat{x}_t),\mathcal{P}(\Delta x_t)\right].$$ - **Tendency Error (TE)** measures their positional agreement via averaged $\ell_1$ distance: $$\mathrm{TE}=\frac{1}{T-1}\sum_{t=1}^{T-1}\mathbb{E}_{ij}\left[\left|\Delta\hat{x}_{t,i,j}-\Delta x_{t,i,j}\right|_1\right].$$ Detailed definitions and motivations are provided in the revised manuscript. We also include the alternative noise constructions discussed in [W2.1, Q2.1]. | Noise type | TWD ↓ (×10⁻³) | TE ↓ (×10⁻³) | | --- | :---: | :---: | | IID Gaussian | 5.009 | **19.924** | | Progressive Noise | 3.809 (−23.96%) | 22.843 (+14.65%) | | Mixed Noise | 3.654 (−27.05%) | 20.704 (+3.91%) | | Dynamical Diffusion | 4.229 (−15.57%) | 20.280 (+1.79%) | | DyPro (ours) | **3.633 (−27.47%)** | 20.630 (+3.54%) | DyPro substantially improves the distributional match and achieves the lowest TWD, with a modest increase in pointwise error compared with IID Gaussian. This aligns with the motivation of DyPro: its early-denoising coupling promotes shared structure across frames, while later relaxation allows local growth and decay, a balance reflected in the closer match to the observed distribution of frame-to-frame changes.
AuthorsOpenReview ↗

Response to Reviewer b78C (Part II)

[Q3] Log-probability normalization and stochasticity

Normalization: We acknowledge the inconsistency between the joint density in the mathematical form and the averaging in the algorithm, which is directly inherited from prior work [3,4,5]. While the joint Gaussian log density should be computed by summing over latent dimensions theoretically, in practice, following [3,4,5] we use the averaged form that potentially promotes numerical stability during optimization and no rescaling tricks are applied. As shown below, summation performs similarly to the baseline without RL, whereas averaging improves all four metrics, empirically supporting this choice. We have clarified this distinction in the manuscript.

NormalizationCSI ↑HSS ↑LPIPS ↓SSIM ↑
w/o RL0.31950.40920.14860.7099
Sum (joint)0.31940.40960.14890.7083
Mean (per-dimension, ours)0.32540.41600.14790.7194

Stochasticity: The stochasticity schedule (κt\kappa_t) directly follows MixGRPO [5], which assigns κtn≡0.7\kappa_{t_n}\equiv 0.7 whenever n∈Wℓn\in \mathcal{W}_\ell and κtn=0\kappa_{t_n}=0 elsewhere. We have added clarification in the manuscript.

[Q4] Robustness across seeds

To assess the robustness of RL improvements, we conduct three independent runs with different random seeds. We report the mean, standard deviation, and 95% confidence intervals across runs.

RunCSI ↑HSS ↑LPIPS ↓SSIM ↑
w/o RL0.31950.40920.14860.7099
seed=420.32280.41300.14680.7200
seed=4120.32420.41500.14680.7196
seed=41120.32540.41600.14790.7194
Mean ± Std0.3241 ± 0.00130.4147 ± 0.00150.1472 ± 0.00060.7197 ± 0.0003
95% CI[0.3209, 0.3274][0.4109, 0.4185][0.1456, 0.1487][0.7189, 0.7204]

The results show low variance across random seeds, with all runs consistently outperforming the DyPro baseline across all metrics.

References

[1] Songwei Ge et al. Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models. ICCV, 2023.

[2] Xingzhuo Guo et al. Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models. ICLR, 2025.

[3] Jie Liu et al. Flow-GRPO: Training Flow Matching Models via Online RL. NeurIPS, 2025.

[4] Zeyue Xue et al. DanceGRPO: Unleashing GRPO on Visual Generation. arXiv:2505.07818, 2025.

[5] Junzhe Li et al. MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE. ECCV, 2026.

AuthorsOpenReview ↗

Request of Reviewer's Attention and Feedback

Dear Reviewer,

This is a kind reminder that three days remain in the discussion period. We would appreciate your feedback on whether our response has addressed your concerns.

Following your suggestions, we have:

  • Added matched-average, reversed, and shuffled reward comparisons.
  • Added correlated-noise baselines and temporal evaluations.
  • Clarified log-probability normalization and the stochasticity schedule.
  • Reported three-seed results with uncertainty estimates.

Thanks again for your valuable review. We look forward to your feedback and are happy to clarify any remaining points.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

We would like to sincerely thank all the reviewers for providing insightful reviews and valuable comments. Your reviews are of great importance to us in improving the quality of this work. We have revised our paper and responded to each reviewer with a separate response.

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

Related papers

Discussion (0)

Sign in to comment.