acceptodds
Under review as a conference paper at ICLR 2027

Multi-Resolution Laplacian Diffusion Models for Kilometer-Scale Climate Emulation

Abstract

Generative models have shown promise in emulating kilometer-scale climate simulations at a small fraction of their cost. A common design pairs a coarse-resolution global generator with a diffusion-based super-resolution stage that generates fine-scale fields conditioned on the coarse output. The two stages are trained independently across a single resolution gap and share no multiresolution structure, so they inhabit different representations and nothing constrains them to compose exactly. Weather data make this harder. Each channel distributes its variance across scales differently, with heavy tails spanning orders of magnitude, and any multiresolution construction must be realized on a spherical grid. We introduce SPyD (Spherical Pyramid Diffusion), which formulates coarse generation and super-resolution as a single diffusion over a Laplacian pyramid. The state splits into a coarse band and orthogonal finer bands, each with its own noise schedule, and the bands evolve under a system of probability-flow ODEs coupled through a shared denoiser, which we instantiate with separately trained conditional band models. In this formulation, the conditioning is an exact projection of the diffusing state, so the super-resolution stage trains directly on this projected state and the coarse and fine bands compose exactly under the pyramid operators. We extend the EDM framework to channel-dependent noise schedules, calibrated per band from measured statistics, and we realize the pyramid on the sphere with global wavelet operators on HEALPix. On ICON simulation output at 5 km, a single SPyD model spanning the full gap in one stage (HPX64 to HPX1024) reproduces the spherical power spectra and pixel distributions of temperature and precipitation and extreme-precipitation percentiles through the th, improves on the deployed state of the art for global kilometer-scale emulation on every channel in CRPS at equal conditioning and equal cost, and produces well-calibrated conditional ensembles as measured by rank histograms. SPyD composes directly with pretrained coarse-resolution generators, supporting end-to-end sampling and reanalysis downscaling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion models, generative climate downscaling, Laplacian pyramids, climate emulation, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_oCx3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SPyD, a multiresolution diffusion framework for kilometer-scale climate emulation. It decomposes atmospheric fields into orthogonal Laplacian wavelet bands on a spherical grid, with channel-dependent noise schedules and a shared diffusion clock. The framework supports both single-stage and multistage super-resolution. Experiments on ICON simulations show improvements over cBottle-SR in CRPS, precipitation extremes, and spatial structure, with additional ensemble calibration and full-pipeline evaluations.

Critical strengths

The use of a shared diffusion framework across resolution bands is an interesting approach to climate downscaling. The channel-dependent noise schedules are particularly relevant because atmospheric variables have very different spatial variability and distributions. The spherical wavelet construction also makes the approach suitable for global applications. The evaluation is reasonably comprehensive, covering 12 atmospheric variables, extreme precipitation, spectral properties, CRPS, and ensemble calibration. Single-stage SPyD improves CRPS over cBottle-SR across all 12 channels, with statistically significant improvements on eight. The precipitation extremes are particularly encouraging, with the 99.999th percentile reproduced within approximately 1.6% of the ICON reference. The matched computational-cost comparison further strengthens the results.

Critical weaknesses

Although the framework is motivated by consistent multiresolution generation, the multistage model performs worse than the single-stage model on several important metrics. In particular, the 99.999th precipitation percentile is underestimated by approximately 16%, and the full generative pipeline produces 20–22% spurious convective cells. The authors acknowledge this issue, which is appreciated, but it limits the practical advantage of the proposed multistage formulation. Exact composition of the resolution bands does not necessarily guarantee accurate generation across stages.

Questions for authors
  1. The multistage model introduces substantial spurious precipitation under generated conditioning. Is this mainly caused by the mismatch between training and generated conditioning, or by errors accumulated across the resolution stages?

  2. How sensitive are the results to the channel-dependent noise calibration? Have the authors compared against a common noise schedule across channels?

  3. The single-stage model performs better than the multistage model in several cases. What would be the main practical motivation for choosing the multistage configuration, apart from supporting intermediate resolutions?

5 comments
AuthorsOpenReview ↗

Response to Reviewer oCx3 (Part 1/3)

We thank the reviewer for the positive and thoughtful evaluation, and for finding no substantive concerns about the soundness of the paper. We appreciate the recognition that a shared diffusion framework across resolution bands is "an interesting approach to climate downscaling", that the channel-dependent noise schedules are "particularly relevant because atmospheric variables have very different spatial variability and distributions", and that the spherical wavelet construction makes the approach "suitable for global applications". We are also encouraged that the reviewer finds the evaluation "reasonably comprehensive", the precipitation extremes "particularly encouraging", and that "the matched computational-cost comparison further strengthens the results". Below, we address the concern about the multi-stage model and answer the three questions.

W1a. The multi-stage model performs worse than the single-stage model on several important metrics.

The multi-stage design follows from how kilometer-scale climate data exist. Simulations are released as a hierarchy of resolutions, one store per HEALPix level [1], and many uses need the intermediate levels rather than only the finest one. The reviewer's concern prompted us to revisit the multi-stage model during the rebuttal, and the revised model is substantially better than the published one and now competitive with the single-stage model.

The published multi-stage model was trained with a much smaller budget than the single-stage model. Its fine stage was a short fine-tune of 17.5k steps, against 126k steps for the single-stage model (Appendix, "Hardware and training cost"). Because the bands compose exactly, we could also measure each stage's contribution separately. This showed that the precipitation gap entered at the intermediate stage, whose output carried only 0.69 of ICON's mesoscale precipitation power. The fine stage generates only scales finer than HPX256, so it cannot recover content missing from the band below. We therefore retrained the multi-stage model with the same training budget as the single-stage model. Its intermediate stage now generates natively at HPX256, and the handoff between stages uses the same wavelet projection as in training. Table A compares the published and revised multi-stage models on the paper's own metrics, with the published single-stage model for reference.

Table A. CRPS per channel (12 held-out months × 16 patches × 8 members; lower is better) and extreme precipitation percentiles (bias against ICON). Channel names as in Table 2; the published columns are copied from Tables 1 and 2 of the paper; bold marks the best of the three.

Single-stageMulti-stageMulti-stage (revised)
cllvi0.043020.049370.04285
clivi0.0041210.0044280.004052
tas0.26940.26710.2576
uas0.49120.48460.4855
vas0.50720.50120.5016
rlut5.5945.6865.451
rsut13.0113.9012.79
pres_msl9.3879.0629.419
pr2.856e-53.921e-53.057e-5
rsds15.3916.7215.32
sst0.11730.11530.1111
sic0.0011320.0012590.001338
pr P99.99 bias−1.5%−11%−8.0%
pr P99.999 bias−1.6%−16%−9.4%

The revised model improves on the published multi-stage model where it was weakest. Precipitation CRPS drops by 22%, the cloud and radiation channels (cllvi, clivi, rlut, rsut, rsds) improve by 4–13%, and the deficit at the 99.999th precipitation percentile falls from −16% to −9.4%. Relative to the single-stage model, it has lower CRPS on 9 of 12 channels, with the largest gains on sst (−5.3%) and tas (−4.4%). It also has the lowest ensemble-mean RMSE among the generative models on 9 of 12 channels. The gap the reviewer identified was therefore a matter of how the published stages were trained, not a limit of the multi-resolution formulation.

W1b. Exact composition of the resolution bands does not necessarily guarantee accurate generation across stages.

Exact composition is a guarantee about the operators, and this is the scope of the claim in the paper. The coarse and generated bands sum exactly to the full field, without loss or overlap at the stage boundary, and the paper reports the measured operator residuals. How accurately each stage generates its band is a learning question, as in any cascade [2]. Exact composition is nevertheless what made the diagnosis above possible: because the bands are orthogonal and sum exactly, the deficit could be attributed to a single stage and fixed there. We will state this distinction explicitly in Section 3.5.

AuthorsOpenReview ↗

Response to Reviewer oCx3 (Part 2/3)

Q1. Is the spurious precipitation caused by the training/generated conditioning mismatch, or by errors accumulated across stages?

The spurious precipitation comes from the mismatch between the conditions the fine stage is trained on and the generated inputs it receives, not from errors accumulating across stages. Under generated conditioning, the intermediate stage's output has a low spurious-cell rate (0.62%), and the fine stage alone raises it to 20–22% (Sec. 4.7). If errors were accumulating across stages, the rate would already be elevated after the intermediate stage. The increase appears only when the fine stage receives inputs unlike those it was trained on.

This is exactly the gap that SPyD's matched noisy conditioning is designed to close. To quantify the effect, we retrained the single-stage model without it, conditioning on clean coarse fields as a standard cascade does. Under generated coarse inputs, this variant loses 14–17% of the fine-scale power that SPyD retains (temperature 0.83× [0.81, 0.84], precipitation 0.86× [0.84, 0.88], 95% paired bootstrap). Matching the conditioning noise is therefore the lever that controls robustness to generated inputs. We also measured the spurious-cell rate directly for the revised multi-stage model, using the same 16 generated coarse samples for all models. Its fine stage no longer increases the rate (0.7× the rate of its intermediate output, against about 35× for the published model), and its final rate matches the single-stage model's (ratio 1.06 [0.99, 1.14], 95% paired bootstrap).

In both the published and the revised multi-stage models, the fine stage is trained on intermediate bands computed from ICON, with matched noise. A natural extension is to train it on outputs of the intermediate stage itself, so that its training and deployment inputs coincide; the shared diffusion clock makes this well-defined because it specifies the noise level at the handoff. This requires generating intermediate-stage samples for the full training set, and we leave it to future work.

Q2. How sensitive are the results to the channel-dependent noise calibration? Have the authors compared against a common schedule across channels?

We appreciate this question, and we ran exactly this comparison. In SPyD, each channel and band has its own noise rate, which sets how quickly noise overwhelms that band as diffusion time increases. The rates are calibrated from the measured band statistics of each channel (Sec. 3.3), so that channels with very different variability lose their signal at comparable points in the diffusion. A common schedule replaces these with one rate shared by all channels, as in image diffusion models. We retrained the single-stage model with one shared rate for all 12 channels, set to the largest calibrated value and keeping everything else identical. We then compared CRPS per channel on the held-out months. If calibration matters, the shared-rate model should be worse on the channels whose statistics differ most from the rest.

That is what we find. With a shared rate, CRPS is significantly worse for precipitation (+1.3% [0.6, 2.1]) and SST (+2.1% [1.2, 3.1]), the heavy-tailed and front-dominated channels, and under generated coarse inputs it also loses 9% of fine-scale temperature power. On smooth, large-scale channels such as pressure and winds, a shared rate performs comparably. The result does confirm your intuition that calibration matters most where atmospheric variables differ most in variability and distribution. It also removes the need to hand-tune a rate per channel. We will add this comparison to the paper.

AuthorsOpenReview ↗

Response to Reviewer oCx3 (Part 3/3)

Q3. What is the main practical motivation for the multi-stage configuration, apart from supporting intermediate resolutions?

Beyond intermediate resolutions, the multi-stage configuration has three practical motivations, at only 8–17% additional inference cost (Sec. 4.1). First, it matches how kilometer-scale data are produced and shared. Simulations are released as one store per HEALPix level [1], and coarse and intermediate levels are far more widely available than the finest one. A fine stage can therefore downscale existing intermediate-resolution products directly, which a model trained for one fixed resolution jump cannot use. Second, the intermediate stage trains and samples on 16× fewer pixels. Third, because the bands compose exactly, errors can be attributed to a single stage; this is how we located and fixed the multi-stage gap during this rebuttal. Exact composition also lets us test each stage in isolation. Given the true HPX256 band of the Figure-2 state as input, the fine stage alone reproduces the typhoon's central pressure within 1 hPa of ICON (912 vs. 911 hPa) and has 2.2–2.9× lower CRPS than every model conditioned on HPX64 alone. The fine stage is therefore already accurate, and further gains will come from the intermediate stage, which the multi-stage framework allows to be trained and improved independently of the fine stage. With the revised model, the multi-stage configuration is now also competitive in accuracy (Table A).

It is also well suited to characterizing extreme weather such as tropical cyclones. In response to Reviewer uFm2, we added storm-scale scores for the cases in Figure 2 and for every strong tropical cyclone (below 985 hPa) in the held-out states; we refer to that response for the protocol. On the Figure-2 typhoon (8-member ensembles), the revised multi-stage model has the central pressure closest to ICON of all models, closer than the single-stage model (917 hPa; ICON 911, single-stage 921, cBottle-SR 938) and keeps the eye clear (eye-to-eyewall rain ratio 0.13, against 0.60 for cBottle-SR). Across the 11 strong held-out cyclones, its radial rain profile is closer to ICON than cBottle-SR's in 10, a significant difference (median −0.75 [−3.30, −0.22] mm/hr). The fractions skill score of its precipitation pattern is the highest of all models in the tropical-cyclone, Maritime Continent and ITCZ regimes (e.g. ITCZ 0.66, against 0.58 for single-stage and 0.37 for cBottle-SR).

We hope our responses have addressed the reviewer's concerns. If any concern remains, please let us know, and we will be happy to respond. Otherwise, we would be grateful if the reviewer would consider whether the new results warrant a more positive assessment.

References

[1] Segura, Hans, et al. "nextGEMS: entering the era of kilometer-scale Earth system modeling." Geoscientific Model Development 18.20 (2025): 7735-7761.

[2] Ho, Jonathan, et al. "Cascaded diffusion models for high fidelity image generation." Journal of Machine Learning Research 23.47 (2022): 1-33.

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

This paper proposes SPyD (Spherical Pyramid Diffusion), a method for generating high-resolution global climate data using multi-resolution diffusion models. It decomposes climate fields into orthogonal resolution bands using wavelet transforms and applies channel- and band-dependent noise schedules. Additionally, it introduces a shared diffusion clock to coordinate coarse-to-fine generation. Experiments on ICON climate simulation data show that single-stage SPyD outperforms the cBottle-SR baseline in probabilistic prediction accuracy, precipitation extremes, and spatial statistics. However, the multi-stage variant performed poorly in many of the evaluations.

Critical strengths
  1. The proposed single-stage SPyD outperformed the cBottle-SR baseline on all 12 climate variables in terms of CRPS. It also achieves more accurate precipitation extremes and spatial statistics, demonstrating its effectiveness for high-resolution climate emulation.
  2. The authors evaluate the approach through multiple metrics, including CRPS, RMSE, precipitation extremes, power spectra, and ensemble calibration. The experiments also consider generated coarse-resolution inputs, which provide useful evidence on how well the method will perform within a complete generative process.
Critical weaknesses
  1. The paper introduces several components to the method, including wavelet-based decomposition, channel-dependent noise schedules, and shared diffusion conditioning. However, there were not enough ablation studies to isolate their individual contributions. Therefore, it’s hard to tell which components contribute to the improved performance compared with the baseline.
  2. Although the paper emphasizes a unified multi-resolution diffusion framework, the multi-stage SPyD generally performs worse than the single-stage version, particularly for precipitation extremes and under generated conditioning. So, I believe there are concerns about if the proposed multi-stage method will be useful and reliable in practice.
Questions for authors
  1. In Sections 3.2–3.4, the authors introduce several components, including wavelet-based decomposition, channel-dependent noise schedules, and noisy conditioning with a shared diffusion clock. However, the experiments in Section 4 mainly compare the complete SPyD framework against cBottle-SR. Could the authors clarify the individual contributions of these components?
  2. In Table 1, multi-stage SPyD performs worse than single-stage SPyD on extreme precipitation percentiles, even when using coarsened ICON fields as conditioning. Furthermore, Appendix F shows that the fine stage substantially amplifies spurious precipitation under generated conditioning. While the authors discuss the sensitivity to generated conditioning and the mismatch between wavelet bases, could they further clarify why multi-stage SPyD also underperforms ground truth coarse conditioning? How do these limitations affect the practical benefits of the proposed multi-stage framework?
  3. In Section 2, the authors discuss prior work such as Edify (which investigated band-dependent Laplacian diffusion) and cascaded diffusion models (which use noisy conditioning). Could the authors clarify what distinguishes SPyD from these existing approaches in terms of methodological novelty, beyond its adaptation to spherical climate data?
5 comments
AuthorsOpenReview ↗

Response to Reviewer DVnP (Part 1/3)

We thank the reviewer for the thoughtful evaluation and for highlighting the core strengths of our work. We are glad that the reviewer values our evaluation "through multiple metrics, including CRPS, RMSE, precipitation extremes, power spectra, and ensemble calibration," including generated coarse inputs, which "provide useful evidence on how well the method will perform within a complete generative process." Below, we address the request for ablations, the multi-stage results, and the question of novelty.

W1 / Q1. The paper introduces several components, including wavelet-based decomposition, channel-dependent noise schedules, and noisy conditioning with a shared diffusion clock. Could the authors clarify their individual contributions?

We thank the reviewer for this question, and we now provide controlled ablations for each component. We retrained the single-stage model from scratch under an identical recipe (same data, backbone, batch size, optimizer, schedule and number of steps), changing exactly one component per retrain. The shared-rate variant replaces the calibrated per-channel noise rates with one rate for all 12 channels, as in Edify [1]. The clean-conditioning variant removes the conditioning noise, so the coarse input carries no noise at any diffusion time, as in a standard cascade. We evaluate each variant in two settings: on true coarse inputs, with the paper's CRPS protocol (12 held-out months × 16 patches × 8 members), and on generated coarse inputs from the cBottle coarse model, the deployment setting of Section 4.7. For the wavelet basis, we compared wavelet families directly at the operator level on the full sphere.

Table A. Component ablations (ratio to SPyD; 95% paired bootstrap).

VariantTruth-conditioned CRPSFine-scale power under generated conditioning
Shared rate (no channel calibration)precipitation 1.013 [1.006, 1.020], SST 1.021 [1.012, 1.030], sea ice 1.037 [0.999, 1.084]temperature 0.91 [0.90, 0.93]
Clean conditioning (no matched noise)within ~1% on all channelstemperature 0.83 [0.81, 0.84], precipitation 0.86 [0.84, 0.88]

Table B. Wavelet families (Section 3.2; operator level on the full sphere, 6→10 split).

WaveletPassband retentionLeakage across the band edgeCoarse-band aliasing (temperature)
Haar (db1)0.8912.0e-3 (17× db4)1.7e-2 (3.7× db4)
db20.9737.0e-47.5e-3
db4 (ours)0.9921.2e-44.6e-3
db60.9952.6e-54.0e-3

We found that, through these ablations, each component performs its intended function. The channel-dependent rates matter most where channel statistics differ most: with one shared rate, precipitation and SST, the heavy-tailed and front-dominated channels become significantly worse, while smooth large-scale channels are served adequately by either. The noisy conditioning matters in the setting it targets: on exact coarse inputs, which a deployed pipeline never sees, clean conditioning is within ~1% CRPS, but on generated inputs it loses 14–17% of the fine-scale power that SPyD retains. The db4 basis keeps the frequency split clean, whereas Haar retains only 89% of the passband and leaks 17× more power across the band edge. db4 is the shortest filter with near-complete passband retention; longer filters bring diminishing returns at the cost of a wider face halo. We will add these results.

W2 / Q2. The multi-stage SPyD performs worse than the single-stage version, including with ground-truth coarse conditioning. Why, and how do these limitations affect its practical benefits?

The multi-stage design follows from how kilometer-scale climate data exist: simulations are released as a hierarchy of resolutions, one store per HEALPix level [2]. The reviewer's concern prompted us to revisit the multi-stage model during the rebuttal, and the revised model is substantially better than the published one and now competitive with the single-stage model.

Why it underperformed with true coarse conditioning. The published multi-stage model was trained with a much smaller budget than the single-stage model: its fine stage was a short fine-tune of 17.5k steps, against 126k steps for the single-stage model (Appendix, "Hardware and training cost"). Because the bands compose exactly, we could measure each stage's contribution separately. This showed that the precipitation gap entered at the intermediate stage, whose output carried only 0.69 of ICON's mesoscale precipitation power. The fine stage generates only scales finer than HPX256, so it cannot recover content missing from the band below, and precipitation extremes are organized at exactly those scales. We therefore retrained the multi-stage model with the same budget as the single-stage model. Its intermediate stage now generates natively at HPX256, and the handoff between stages uses the same wavelet projection as in training.

AuthorsOpenReview ↗

Response to Reviewer DVnP (Part 2/3)

Table C. CRPS per channel (12 held-out months × 16 patches × 8 members; lower is better) and extreme precipitation percentiles (bias against ICON). Channel names as in Table 2; the first two columns are copied from Tables 1 and 2 of the paper; bold marks the best of the three.

Single-stageMulti-stageMulti-stage (revised)
cllvi0.043020.049370.04285
clivi0.0041210.0044280.004052
tas0.26940.26710.2576
uas0.49120.48460.4855
vas0.50720.50120.5016
rlut5.5945.6865.451
rsut13.0113.9012.79
pres_msl9.3879.0629.419
pr2.856e-53.921e-53.057e-5
rsds15.3916.7215.32
sst0.11730.11530.1111
sic0.0011320.0012590.001338
pr P99.99 bias−1.5%−11%−8.0%
pr P99.999 bias−1.6%−16%−9.4%

The revised model improves on the published multi-stage model where it was weakest: precipitation CRPS drops by 22%, the cloud and radiation channels improve by 4–13%, and the deficit at the 99.999th precipitation percentile falls from −16% to −9.4%. Relative to the single-stage model, it has lower CRPS on 9 of 12 channels, with the largest gains on sst (−5.3%) and tas (−4.4%), and it has the lowest ensemble-mean RMSE among the generative models on 9 of 12 channels. The gap in the published results was therefore a matter of how the stages were trained, not a limit of the multi-resolution formulation.

Why the fine stage amplified spurious precipitation under generated conditioning (Appendix F). The amplification came from the mismatch between the conditions the fine stage was trained on and the generated inputs it received, not from errors accumulating across stages: the intermediate output had a low spurious-cell rate (0.62%), and the fine stage alone raised it to 20–22%. This is the gap that SPyD's noisy conditioning is designed to close, and the clean-conditioning ablation in Table A shows that it is the relevant lever. We also measured the spurious-cell rate directly for the revised multi-stage model, on the same 16 generated coarse samples for all models. Its fine stage no longer increases the rate (0.7× the rate of its intermediate output, against about 35× for the published model), and its final rate matches the single-stage model's (ratio 1.06 [0.99, 1.14], 95% paired bootstrap).

Practical benefits. The single-stage and multi-stage models are two instances of the same framework, and the choice between them is a deployment trade-off. Beyond intermediate resolutions, the multi-stage configuration has three practical benefits, at only 8–17% additional inference cost (Section 4.1). First, it matches how kilometer-scale data are produced and shared: coarse and intermediate HEALPix levels are far more widely available than the finest one, and a fine stage can downscale intermediate-resolution products directly. Second, the intermediate stage trains and samples on 16× fewer pixels. Third, because the bands compose exactly, errors can be attributed to a single stage; this is how we located and fixed the gap above. Exact composition also lets us test each stage in isolation. Given the true HPX256 band of the Figure-2 state as input, the fine stage alone reproduces the typhoon's central pressure within 1 hPa of ICON (912 vs. 911 hPa) and has 2.2–2.9× lower CRPS than every model conditioned on HPX64 alone. The fine stage is therefore already accurate, and further gains will come from the intermediate stage, which the multi-stage framework allows to be trained and improved independently of the fine stage.

The revised multi-stage model is also well suited to characterizing extreme weather such as tropical cyclones (storm metrics, protocol and tables in our response to Reviewer uFm2). On the Figure 2 typhoon, its central pressure is the closest to ICON of all models, closer than the single-stage model (917 hPa; ICON 911, single-stage 921, cBottle-SR 938), and it keeps the eye clear (eye-to-eyewall rain ratio 0.13, against 0.60 for cBottle-SR). Across the 11 strong held-out tropical cyclones, its radial rain profile is closer to ICON than cBottle-SR's in 10, a significant difference (median −0.75 [−3.30, −0.22] mm/hr). Together with its lower CRPS than the single-stage model on 9 of 12 channels (Table C), this makes the revised multi-stage model a competitive choice for the distributional and extreme-event statistics that downscaling models target.

AuthorsOpenReview ↗

Response to Reviewer DVnP (Part 3/3)

Q3. What distinguishes SPyD from Edify and cascaded diffusion models in terms of methodological novelty, beyond its adaptation to spherical climate data?

We thank the reviewer for this question, which helps us sharpen the contribution statement. Reviewer uFm2 also finds the method "interesting and novel", and Reviewer oCx3 notes that the channel-dependent schedules are "particularly relevant because atmospheric variables have very different spatial variability and distributions". Beyond the spherical setting, SPyD differs from both lines of work in three methodological respects, each quantified by an ablation in Table A or B.

1. Calibrated per-channel, per-band noise rates. Edify [1] diffuses Laplacian bands of RGB images with band-dependent noise rates that are set per frequency band and shared across the three color channels, which suits images whose channels have similar statistics. SPyD targets 12 physical fields that differ by orders of magnitude in variance, spectral slope and tail weight. It derives a rate for every channel and band from measured band statistics, chained across levels by a recursive handoff rule (Section 3.3), so no per-channel tuning is needed. Replacing it with one Edify-style shared rate makes precipitation and SST CRPS significantly worse (Table A).

2. Conditioning noise set by the generator's own clock. Cascaded diffusion [3] corrupts the low-resolution conditioning with noise augmentation whose level is a hyperparameter, chosen independently of how the coarse model generates and tuned at inference. The fine stage learns to ignore a corruption whose statistics it never sees at deployment. In SPyD, the condition is the exact projection of the diffusing state, noised on the same diffusion clock as the generated band (Section 3.4). The noise the condition carries at the handoff is therefore specified by the framework rather than tuned. Removing it loses 14–17% of fine-scale power under generated inputs (Table A).

3. Orthonormal spherical bands with exact composition. Image pyramids are defined on planar grids. SPyD builds face-halo db4 wavelet operators on the HEALPix sphere whose coarse projector, band projectors and decimating analysis/synthesis pair are mutually consistent: the bands sum exactly to the field without overlap, with measured residuals at face seams and corners (Appendix). This makes band-wise generation on the sphere well-posed, and it is what allowed the stage-wise diagnosis in our answer to W2. Table B quantifies the choice of basis.

Together, these turn the collection of design choices in prior multi-resolution diffusion into one framework with a common clock, in which noise rates, conditioning noise and handoffs are derived rather than tuned. We will revise the contribution list in the introduction to state these three points and point to the corresponding ablations.

We hope our responses have addressed the reviewer's concerns. If any concern remains, please let us know, and we will be happy to respond. If our responses have resolved your concerns, we would be grateful if you would kindly consider raising your score.

References

[1] Atzmon, Yuval, et al. "Edify Image: High-quality image generation with pixel space Laplacian diffusion models." arXiv preprint arXiv:2411.07126 (2024).

[2] Segura, Hans, et al. "nextGEMS: entering the era of kilometer-scale Earth system modeling." Geoscientific Model Development 18.20 (2025): 7735-7761.

[3] Ho, Jonathan, et al. "Cascaded diffusion models for high fidelity image generation." Journal of Machine Learning Research 23.47 (2022): 1-33.

Reviewer_uFm2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SPyD, a diffusion model for turning coarse global climate fields into finer fields at about 5 km resolution. It splits each field into wavelet bands on the HEALPix grid. Each band and weather variable gets a noise level based on its measured data. The authors train a one-stage model from HPX64 to HPX1024 and a two-stage model with HPX256 in between. On held-out ICON simulation data, the one-stage model has lower CRPS than cBottle-SR on all variables. It also matches the pooled upper tail of precipitation more closely. The paper tests both downscaling of true coarse fields and downscaling of fields from a coarse generator.

Critical strengths
  1. Strong one-stage results on several useful measures. The paper compares with cBottle-SR using the same coarse inputs and similar sampling cost. The one-stage model improves CRPS on all 12 variables. Its precipitation spectrum and high percentiles are also closer to ICON.
  2. Careful treatment of the spherical grid. The authors explain how the wavelet bands work on HEALPix. This makes the method easier to assess and reuse.
Critical weaknesses
  1. The comparison with other methods is limited. cBottle-SR is the only external learned super-resolution baseline. Naive upsampling is useful as a reference, but it is not a strong generative method. This makes it hard to tell whether SPyD would also improve on other approaches. Figure 2 suggests better storm and rain structure, but it gives no numerical scores for the cases shown. The pooled scores elsewhere do not directly measure the cyclone eye, rainbands, or local rain patterns highlighted in that figure.
  2. The paper does not isolate which design choices cause the gains. The method changes the wavelet basis, band structure, noise rates, and conditioning noise. Without ablation study, it is hard to tell which parts of SPyD are needed for the one-stage improvement. For example, how can you proof the effectiveness of the channel-dependent noise schedules in the contribution?
Questions for authors

I find the proposed method interesting and novel. Beyond the comparison and ablation concerns above, I have three questions:

  1. In the upper-left panel of Figure 3, the temperature curves show some sharp changes. Could you explain how the curves were computed and share the plotted values or code?
  2. The multi-stage model is less accurate than the single-stage model at the highest precipitation percentiles in Table 1 and on several CRPS scores in Table 2, although it does better on some other variables. How do you interpret these results in light of the claim that the bands compose well? Could you also report the ensemble-mean RMSE for the multi-stage model in Table 2?
  3. The ICON output is remapped to HEALPix for this study. How well would SPyD transfer to other Earth system models or grids? Please briefly explain whether HEALPix is common in this setting and whether remapping could affect fine-scale structure or extremes.
7 comments
AuthorsOpenReview ↗

Response to Reviewer uFm2 (Part 1/5)

We thank the reviewer for the thoughtful evaluation and for highlighting the core strengths of our work. We appreciate the recognition of SPyD's "strong one-stage results on several useful measures", with CRPS improving "on all 12 variables" and the "precipitation spectrum and high percentiles" closer to ICON, of the fairness of the comparison "using the same coarse inputs and similar sampling cost", and of our "careful treatment of the spherical grid", which "makes the method easier to assess and reuse". We are also glad that the reviewer finds "the proposed method interesting and novel".

The reviewer's two weaknesses ask for comparisons beyond cBottle-SR with numerical scores for the cases in Figure 2, and for ablations that isolate which design choices cause the gains. We ran new experiments for both and address W1 with a new regression baseline and storm- and regime-level scores for Figure 2. We also address W2 with controlled ablations of each component. Finally, we answer the three questions, including a revised multi-stage model with its ensemble-mean RMSE.

W1a. The comparison with other methods is limited: cBottle-SR is the only external learned super-resolution baseline.

Why cBottle-SR is the primary baseline. cBottle-SR is the state of the art for global, multivariate, km-scale downscaling [1]. To our knowledge, it is the only released downscaler trained on global km-scale ICON output, and it is trained on exactly the simulation we use. Even cBottle-SR does not include any further baseline comparisons as well. Other released diffusion downscalers, such as CorrDiff [2], are regional (Taiwan, CONUS) and trained on different inputs and simulators, so a comparison against them would measure domain transfer rather than the method. Naive upsampling is included in the paper only as a floor, which shows how little of the tail the coarse input carries (its P99.999 is 87% low).

New baseline: deterministic U-Net regression. To compare against a different class of method, we added a deterministic U-Net regression model, a standard baseline in learned downscaling [2, 3]. It is trained with everything else matched to SPyD: the same SongUNet backbone (~330M parameters), the same training data, 128×128 patches and HPX64 conditioning, and an MSE objective normalized per channel. Its results below are from an interim checkpoint.

Table A. Regression baseline vs. SPyD and cBottle-SR (32 held-out states; CRPS on the paper's 12-month protocol; power relative to ICON; meso = 0.02–0.2 km⁻¹, fine = 0.2–0.8 km⁻¹).

CRPS relative to SPyD (12 channels)Precip. P99.999 vs. ICONPrecip. power, meso / fineTemperature power, fine
U-Net regression1.47–2.14× higher on every channel (precip. 1.87× [1.83, 1.91])−74%0.45 / 0.030.39
cBottle-SR1.00–1.13× (Table 2)+13.8%1.17 / 1.321.00
SPyD—−1.4%0.94 / 0.960.94

Table B. Pooled precipitation percentiles (mm/hr, 32 held-out states, 403M pixels; relative bias against ICON; naive upsampling, cBottle-SR and SPyD from Table 1).

P99P99.9P99.99P99.999
ICON2.2322.4972.58132.16
Naive upsampling2.33 (+4%)5.79 (−74%)10.58 (−85%)16.89 (−87%)
U-Net regression2.44 (+10%)8.84 (−61%)20.02 (−72%)34.51 (−74%)
cBottle-SR2.15 (−4%)25.16 (+12%)85.20 (+17%)151.30 (+14%)
SPyD2.19 (−2%)21.62 (−4%)71.51 (−1.5%)130.03 (−1.6%)

A regression model trained with MSE estimates the conditional mean, so it produces smooth fields: it keeps only 3% of ICON's fine-scale precipitation power and 39% of the fine-scale temperature power, and its precipitation tail is closer to naive upsampling than to ICON, with the 99.999th percentile 74% lower. As in CorrDiff's comparison of deterministic regression and diffusion downscaling [2], and as in our paper, CRPS is the main relevant score for this task, because it evaluates the full predictive distribution and is strictly proper [4]. On CRPS, regression is 1.5–2.1× worse than SPyD on every channel, while SPyD stays within 6% of ICON's power at both scales and within 4% of ICON at every precipitation percentile.

Beyond this baseline, the two ablation variants in W2 compare SPyD against the key design choices of two other approaches under an identical recipe: clean conditioning, as used in standard cascaded diffusion downscalers, and a single noise rate shared across channels, as in Edify's Laplacian diffusion [6]. Unsurprisingly, both perform worse than SPyD (Table F).

AuthorsOpenReview ↗

Response to Reviewer uFm2 (Part 2/5)

W1b. Figure 2 suggests better storm and rain structure...

We thank the reviewer for this suggestion, which makes the claims in Figure 2 quantitative. We added storm-scale metrics for the Figure 2 typhoon and for every strong tropical cyclone in the held-out states, and fractions skill scores for the other Figure 2 regimes.

Tropical cyclones are detected in ICON by central pressure and 10 m wind. For each model, the field is recentered on that model's own pressure minimum, so displacement is not penalized twice. We compute central pressure, maximum 10 m wind, eye clearance (mean precipitation within 20 km relative to the eyewall peak), and the error and correlation of the azimuthal-mean precipitation profile against ICON, which captures eye, eyewall, and rainbands in one curve. Thresholds and metrics were fixed before any cBottle-SR result was computed.

Table C. The Figure 2 typhoon. Each model draws 8 members from the same coarse input; we report the ensemble mean with a 95% bootstrap interval.

Central pressure (hPa)Max 10 m wind (m/s)Eye rain / eyewall peakRain-profile RMSE (mm/hr)Profile correlation
ICON91156.00.01——
cBottle-SR938 [937, 940]47.0 [44.8, 49.4]0.60 [0.39, 0.79]9.3 [8.1, 10.3]0.57 [0.45, 0.71]
SPyD (single-stage)921 [920, 921]59.6 [58.2, 60.9]0.10 [0.08, 0.13]2.8 [2.5, 3.1]0.98 [0.97, 0.99]
SPyD (multi-stage)917 [915, 919]60.6 [59.8, 61.5]0.13 [0.11, 0.16]3.5 [3.0, 4.1]0.95 [0.93, 0.96]

SPyD keeps the eye clear and reproduces the radial rain structure (0.98), while cBottle-SR fills 60% of the eye with eyewall-strength rain (0.57). All differences from cBottle-SR are significant, e.g., the paired profile-RMSE difference is −6.5 [−7.5, −5.3] mm/hr and the eyewall-radius error drops by 25 [18, 33] km. SPyD is also closer in intensity: the multi-stage model's central pressure is the closest to ICON of all models (6 hPa too weak), single-stage 10 hPa, cBottle-SR 27 hPa.

Table D. All strong tropical cyclones (< 985 hPa) in the 32 held-out states (n = 11). Each entry is the median per-storm difference in error against ICON, SPyD minus cBottle-SR, with a 95% bootstrap interval over storms; negative means SPyD is closer.

Rain-profile RMSE (mm/hr)Storms where SPyD is closer
SPyD single-stage vs. cBottle-SR−0.57 [−2.39, −0.04]10 / 11
SPyD multi-stage vs. cBottle-SR−0.75 [−3.30, −0.22]10 / 11

The median eye-to-eyewall rain ratio is 0.03 in ICON, 0.06 for both SPyD models and 0.22 for cBottle-SR. The Figure-2 typhoon is therefore representative: SPyD's advantage in eye and eyewall structure holds across the strong held-out cyclones.

Table E. Fractions skill score (FSS) [5] for the other Figure-2 regimes. Scores are for the precipitation pattern in 25 km neighborhoods and thresholds of 1/5 mm/hr, averaged over 8 members; higher is better.

RegimecBottle-SRSPyD
Tropical cyclone0.76 / 0.670.82 / 0.80
Extratropical cyclone0.71 / 0.670.75 / 0.75
Maritime Continent convection0.51 / 0.250.57 / 0.27
ITCZ0.37 / 0.300.58 / 0.46
Southern Ocean front0.930.91

FSS measures whether rain falls in the right places at a given scale, which pooled statistics cannot. SPyD scores higher in four regimes, with the largest gain in the ITCZ (0.58 [0.53, 0.64] vs. 0.37 [0.29, 0.44] at 1 mm/hr). At the Southern Ocean front, both models score above 0.9; no pixel in that box exceeds 5 mm/hr, so there is no 5 mm/hr score. We will add Tables C–E to the paper next to Figure 2.

W2. The paper does not isolate which design choices cause the gains. How can the effectiveness of the channel-dependent noise schedules be shown?

We thank the reviewer for this question, and we now provide controlled ablations for each component.

Ablating, we retrained the single-stage model from scratch using the same recipe (same data, backbone, batch size, optimizer, cosine schedule, and number of steps), changing exactly one component per retraining. The shared-rate variant replaces the calibrated per-channel band noise rates with a single rate across all 12 channels, set to the largest calibrated value, ensuring that every channel remains noise-dominated at the top of the diffusion window; per-channel preconditioning is retained, so only the calibration is removed. This is the Edify-style choice [6]. The clean-conditioning variant removes the matched conditioning noise, so the coarse condition carries no noise at any diffusion time, as in a standard cascade, at both training and inference. All variants are evaluated at the same checkpoint, on identical patches for truth-conditioned CRPS, and on cBottle coarse samples for generated conditioning, the deployment setting of Section 4.7.

AuthorsOpenReview ↗

Response to Reviewer uFm2 (Part 3/5)

Table F. Component ablations (ratio to SPyD; 95% paired bootstrap).

VariantTruth-conditioned CRPSFine-scale power under generated conditioning
Shared rate (no channel calibration)precipitation 1.013 [1.006, 1.020], SST 1.021 [1.012, 1.030], sea ice 1.037 [0.999, 1.084]temperature 0.91 [0.90, 0.93]
Clean conditioning (no matched noise)within ~1% on all channelstemperature 0.83 [0.81, 0.84], precipitation 0.86 [0.84, 0.88]

Channel-dependent noise schedules (the reviewer's example). One shared rate significantly worsens CRPS on precipitation and SST, the channels whose band statistics are furthest from the rest: heavy-tailed, intermittent precipitation, and SST with sharp fronts and a land mask. A single rate necessarily under-noises some channels and over-noises others relative to their band variance, so the noise level at which each channel's fine-scale structure is decided differs across channels, and the heavy-tailed channels suffer most. Under generated conditioning, the shared rate also loses 9% of fine-scale temperature power. On smooth, large-scale channels such as pressure and winds, a shared rate is adequate, as one would expect. The calibration matters where channels differ most, and it removes the need to hand-tune rates per channel. We will state this scope explicitly in Section 3.3.

Matched noisy conditioning. On exact coarse inputs, the clean-conditioning variant is within ~1% CRPS, since it is trained on exactly those inputs. A deployed pipeline never sees exact coarse inputs; it sees generated ones, which is the setting the stage handoff exists for. There, clean conditioning loses 14–17% of the fine-scale power that SPyD retains, which is precisely the effect the shared clock is designed to produce.

Wavelet basis. We compared wavelet families on the full sphere at the operator level, for the 6→10 split.

Table G. Wavelet families.

WaveletPassband retentionLeakage across the band edgeCoarse-band aliasing (temperature)
Haar (db1)0.8912.0e-3 (17× db4)1.7e-2 (3.7× db4)
db20.9737.0e-47.5e-3
db4 (ours)0.9921.2e-44.6e-3
db60.9952.6e-54.0e-3

Haar, the simplest choice, retains only 89% of the passband, leaks 17× more power across the band edge than db4, and puts 3.7× more blocky aliasing into the coarse band. db4 is the shortest filter with near-complete passband retention; longer filters bring diminishing returns at the cost of a wider face halo. Single- versus multi-stage isolates the band structure (Q2). We will add Tables F and G to the paper.

Response to the Questions

Q1. In Figure 3, the temperature curves show some sharp changes. How were the curves computed, and can you share the plotted values or code?

The distributions are histograms of all pixels pooled over the 32 held-out states (403M values per model, 200 bins), and the spectra are spherical-harmonic power spectra computed on the full HPX1024 sphere and averaged over the same states. The sharp features are present in the ICON reference itself and are physical. The peak at 273 K comes from melting snow and ice surfaces, which are pinned at the melting point. The fourfold drop in pixel count between 300 and 301 K is the warm-ocean ceiling: over tropical oceans, 2 m temperature tracks the sea surface, which rarely exceeds about 301 K. The steps at the extremes of the range are the last few pixels of a 403M-pixel sample. We will release the plotting script and the plotted values with the supplementary material.

Q2. How do you interpret the multi-stage results in light of the claim that the bands compose well? Could you also report the ensemble-mean RMSE for the multi-stage model in Table 2?

Exact composition is a guarantee about the operators, and this is the scope of the claim in the paper. The coarse and generated bands sum exactly to the full field, without loss or overlap at the stage boundary, and the paper reports the measured operator residuals. How accurately each stage generates its band is a learning question, as in any cascade [7]. Exact composition is nevertheless what allowed us to diagnose the multi-stage results: because the bands are orthogonal and sum exactly, each stage's contribution can be measured separately. We will state this distinction explicitly in Section 3.5.

Comments on the paper

1 comment

Dear Area Chair, Program Chairs, and Reviewers,

As the discussion concludes, we thank the reviewers for their constructive feedback and the Area Chair for confirming that the full rebuttal will be considered. We summarize SPyD's contributions and new evidence below.

SPyD (Spherical Pyramid Diffusion) introduces a unified multi-resolution diffusion framework for kilometer-scale global climate emulation. Unlike conventional pipelines that train coarse generators and super-resolution models independently across a single resolution gap, SPyD decomposes global fields into orthogonal spherical wavelet bands, calibrates channel- and band-dependent noise schedules from climate statistics, and couples generation through a shared diffusion clock. Projected conditioning and exact band composition enable consistent single- and multi-stage downscaling on HEALPix.

Key results:

  • Probabilistic accuracy: Single-stage SPyD beats cBottle-SR on CRPS across all 12 variables at matched conditioning and cost (significant on eight).
  • Extremes: The precipitation P99.999 is within 1.6% of ICON, versus a 14% overestimate by cBottle-SR.
  • Storm structure (new): For the Figure 2 typhoon, radial precipitation-profile correlation improves from 0.57 (cBottle-SR) to 0.98; SPyD has lower profile error for 10 of 11 strong held-out cyclones.
  • Multi-stage revision (new): The revised model improves precipitation CRPS by 22% over the original multi-stage model, beats single-stage CRPS on 9 of 12 variables, and no longer amplifies spurious precipitation under generated conditioning.

Reviewer uFm2 described SPyD as "interesting and novel" and praised its strong one-stage results and spherical-grid treatment. In response to concerns about baselines, missing ablations, and storm metrics, we added a matched U-Net regression baseline (1.47–2.14× worse CRPS than SPyD), controlled component ablations, and quantitative cyclone and precipitation-pattern evaluations. The reviewer explicitly stated that "the matched U-Net baseline and the ablation results address most of my concerns", agreed that cBottle-SR is the closest directly comparable released model, and raised their rating from 2 to 3.

Reviewer DVnP recognized improvements in CRPS, precipitation extremes, and spatial statistics, but questioned component contributions, multi-stage reliability, and novelty relative to Edify and cascaded diffusion.

  • Ablations: A shared noise schedule worsens precipitation CRPS by 1.3% and SST by 2.1%; removing matched noisy conditioning loses 14–17% of fine-scale power under generated inputs. Wavelet comparisons quantify the benefits of db4.
  • Multi-stage: We traced the original precipitation deficit to the intermediate stage and retrained with a matched budget and consistent wavelet handoffs. The revised model achieves lower CRPS than single-stage on 9 of 12 channels and eliminates spurious-precipitation amplification.
  • Novelty: We clarified three methodological contributions: statistically calibrated per-channel/per-band noise rates, conditioning determined by the generator's shared diffusion clock, and exactly composable spherical wavelet operators, beyond a spherical adaptation of existing methods.

Importantly, uFm2 raised overlapping concerns about missing ablations and multi-stage performance, acknowledged that the new evidence addressed most concerns, and raised their rating from 2 to 3. We received no response from DVnP and therefore could not establish whether the same evidence resolved their concerns.

Reviewer oCx3 (3) called the shared framework "an interesting approach to climate downscaling" and the channel-dependent schedules "particularly relevant." Their main concerns were multi-stage extreme-precipitation deficits and spurious convective cells. The revised model reduces the P99.999 deficit from −16% to −9.4%; its fine stage no longer amplifies spurious cells (versus ~35× previously), and its final spurious-cell rate matches single-stage within uncertainty. The multi-stage model also predicts the Figure 2 typhoon's central pressure within 6 hPa of ICON, versus 27 hPa for cBottle-SR. We additionally supplied the requested shared-schedule ablation and explained the benefits of intermediate-resolution stages and stage-wise diagnosis.

Reviewer uFm2 engaged and increased their score; DVnP and oCx3 did not respond, despite our rebuttals and follow-up reminders. We hope particular consideration will be given to the controlled ablations and revised multi-stage results, which directly address the central concerns underlying DVnP's rating of 2.

We will incorporate the new experiments, clarifications, and acknowledged limitations into the revised manuscript. Thank you for your consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.