Review for the GMD

This manuscript presents a timely and meaningful intercomparison of three “most used” gen AI models for precipitation super-resolution downscaling. The study addresses an important problem at the interface of atmospheric science and machine learning, and the effort to compare multiple model classes within a unified framework is valuable. At the same time, several aspects of the manuscript require substantial clarification and strengthening before the conclusions can be fully supported. Addressing these issues would improve the scientific rigor and impact of the paper.

Major Comments : 

1.The authors state that the 10-member ensemble is generated from 10 independently trained models initialized with different random seeds. This procedure primarily reflects epistemic uncertainty associated with parameter estimation and training variability. However, the central theoretical advantage of conditional generative models is that for a given low-resolution input, they can generate a distribution of plausible high-resolution outputs through stochastic sampling. At present, the manuscript uses one prediction from each independently trained model and interprets the resulting spread as ensemble uncertainty, which is not equivalent to sampling the conditional output distribution of a single trained generative model. The authors must separate these two uncertainty sources explicitly. In addition to the current analysis, they should report results from repeated stochastic sampling using a single trained model, preferably the best-performing checkpoint, and compare that spread with the spread arising from different training seeds. This distinction is essential for a correct interpretation of the ensemble results.

2.The manuscript refers in several places to ERA5-Land as “observation”. This terminology is not correct. ERA5-Land is a reanalysis-based product, not a direct observational dataset. Since the study does not use in situ station obs, radar, satellite retrievals, or soundings as reference truth, the manuscript should consistently refer to ERA5-Land as a reanalysis or reanalysis-based target, not as observation. You could read this paper to obtain the detailed reason.

https://doi.org/10.1175/BAMS-D-14-00226.1

3.The use of min-max normalization may help stabilize training, but it raises an important concern for precipitation, especially for extreme events. Min-max scaling bounds the normalized target by the range seen in the training data, which may hinder robust extrapolation to unprecedented values. This issue is especially relevant for climate-related downscaling and extreme precipitation, where out-of-sample events may exceed the historical training maximum. This issue may also be relevant to the behavior shown in Figure S11, where DDPM with T=100 approaches the upper bound and cannot grow freely. The authors should discuss this limitation explicitly and test at least one alternative normalization strategy such as quantile normalization or z-score normalization over wet pixels only, reporting whether the normalization choice materially changes the extreme-value results.

4.Precipitation is not a typical continuous variable like temperature, pressure, or geopotential height. It is sparse, intermittent, highly skewed, and often better represented by zero-inflated or Tweedie-like distributions. For this reason, the loss-function choice deserves much more discussion than it currently receives. The authors should discuss why their selected losses are appropriate for precipitation specifically, and whether distribution-aware losses could improve tail behavior and wet-day occurrence. Recent studies suggest that distributional losses can be beneficial for precipitation prediction and downscaling. At minimum, this should be discussed more clearly. Ideally, the authors would include a sensitivity test or ablation experiment.

https://arxiv.org/html/2509.08369

https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2024GL111828

5.Figure 5 appears to compare model predictions against an upsampled low-resolution field rather than the native high-resolution ERA5-Land target, given the clustering of identical reference values on the x-axis. If so, the comparison is not appropriate and the figure needs to be redone using the actual high-resolution target field. If this interpretation is incorrect, the authors should clarify exactly how the reference field in Figure 5 was constructed.

6.The spatial lag analysis in Fig. 6 is not the most informative way to evaluate scale-dependent structure for precipitation super-resolution. A spatial power spectrum would be more standard and more physically interpretable. The authors should add a spectral analysis to the main paper. The current spatial lag figure could be moved to the supplement.

7.All three models condition only on coarse-resolution precipitation. Precipitation is not a self-contained variable. For instance, topography is a key control on high-resolution precipitation structure, especially in regions where orographic effects are important. The manuscript does not sufficiently discuss the implications of omitting terrain height or other static geographic information as conditioning variables. The smoothness seen in the deterministic baseline may partly reflect the lack of physically informative conditioning, rather than only the architecture itself. This point is also relevant for the generative models. One of the strengths of conditional DDPM is the flexibility with which conditioning information can be incorporated, including modulation-based conditioning (like FiLM used here and AdaGN etc.). The authors should discuss more directly whether including terrain or other physically meaningful covariates could materially change the conclusions.

8.A plain UNet trained with a pointwise loss is well known to produce overly smooth outputs in super resolution tasks, so the observed contrast between UNet and generative models may partly reflect the baseline choice rather than an inherent limitation of deterministic approaches. The paper should justify this baseline more carefully or include at least one stronger deterministic baseline such as a sub-pixel convolution or PixelShuffle-based architecture.

https://arxiv.org/pdf/1609.05158

9.SSIM is a perceptual image metric designed for natural-image comparison based on luminance and contrast, and its physical meaning for sparse, intermittent precipitation fields is limited. SSIM should not be emphasized as a primary result and may be moved to the supplement. The Q-Q diagnostics currently in Figure S7 are more physically meaningful for a heavy-tailed intermittent variable and should be promoted to the main text.

10.Because the low-resolution inputs are constructed using a block-averaging operator, the mass conservation result in Section 5.2.2 is partly guaranteed by the experimental design and is less informative than the manuscript implies. This section should be shortened and some of the discussion moved to the supplement.

11.The manuscript compares a U-Net against generative models, but it does not include a simple interpolation baseline. That omission weakens the benchmark. At minimum, the authors should include one standard interpolation baseline so that readers can assess whether the deterministic neural model actually adds value beyond trivial reconstruction.

12.Precipitation has strong temporal autocorrelation because it is tied to evolving synoptic and mesoscale systems. Wet-spell duration, dry-spell duration, and multi-day persistence are among the most hydrologically relevant properties of any downscaled product. A model may match daily spatial structure while still failing to reproduce realistic persistence across time. The manuscript does not evaluate this sufficiently. The authors should either include temporal diagnostics that directly assess persistence behavior or clearly state that the current evaluation is not enough to establish hydrological usefulness.

Some minor issues : 

1.Figures 3 and S3 should state the timestamp or time period shown. The figure captions should clearly indicate which date or sample is being plotted.

2.At line 126, the manuscript refers to one region as the “Pacific Northwest”. Based on the domain shown, this terminology appears inaccurate, since Utah and western Nevada are not usually considered part of the Pacific Northwest. It would be better to use “Northwest” unless the domain is redefined.

3.The manuscript states that the models are trained on the Central Plains and Northwest, while validation uses the Central Plains plus a subset of the Northeast, and the remaining Northeast samples are used for independent testing. This is understandable, but the exact fractions or sample counts should be stated explicitly in the main text.

4.The manuscript sets daily precipitation below 1 mm per day to zero and excludes days with fewer than 1 percent wet pixels. These choices may be reasonable, but the authors should report how many samples are removed by region and season and briefly discuss the potential impact on light-rain statistics and wet-day occurrence.

5.The manuscript states that the models are trained on the Central Plains and Northwest, while validation uses the Central Plains plus a subset of the Northeast, and the remaining Northeast samples are used for independent testing. This is understandable, but the exact fractions or sample counts should be stated explicitly in the main text.

6.The U-Net training description omits weight decay and the Adam beta values; since the WGAN uses non-default beta1=0.0 and beta2=0.9, any deviation from defaults for other models must also be explicitly reported. The DDPM section does not specify the optimizer, initial learning rate, or weight decay. 

 1. General Comments

The study compares three deep learning frameworks (U-NET, WGAN, and DDPM) for precipitation super-resolution across different hydroclimatic regimes. While the comparison is timely, there are significant critical concerns regarding the experimental design and the technical execution that need to be addressed before the paper can be considered for publication.

2. Major Concerns

• Training Stability and Overfitting (Critical Concern) A fundamental issue exists regarding the training convergence and generalization of the generative models, particularly the DDPM. In Section 3 (Lines 398–400), the authors describe the dataset splitting into training and testing sets, but no validation set is mentioned. However, validation curves are provided in the Supplemental Material (Figure S2), at least for the UNET and the DDPM models. They are missing for the WGAN model. Upon inspection of Figure S2, despite the compressed scale of the y-axis, the DDPM model appears to exhibit clear signs of overfitting after approximately epoch 10. The divergence between training and validation loss suggests that the model is no longer learning generalizable features of the precipitation fields but is instead memorizing the training samples. Since the remainder of the paper relies on the results derived from these trained weights, the validity of the inter-comparison and the subsequent conclusions regarding DDPM’s performance are in question. Also, the training curves of the WGAN raise some concerns, and no validation curve is shown. The authors must:
  • Clarify the lack of a validation set description in the main text and for WGAN in FigureS2
  • Provide a detailed analysis of the loss curves with a more appropriate y-axis scale. Is the overfitting actually happening?
  • Address how they ensured the "optimal" stopping point for training to prevent reporting results from an overfit model.
  •  In lines 430-434, the authors acknowledge the potential for overfitting but treat it in a too simplistic manner.
• Directly related to the previous point: Sample Size vs. Model Complexity (Critical)

  The technical rigor of the study is challenged by the potential imbalance between the available data and the model's capacity. The authors utilize ~33,000 daily fields (with only ~19,200 samples for training) to train high-capacity architectures. Modern WGAN and DDPM implementations (including features mentioned like sinusoidal time embeddings and FiLM layers) often contain millions of trainable parameters.

  • The authors must provide a table explicitly summarizing the total number of trainable parameters for the U-NET, WGAN, and DDPM.
  • Given the relatively small training set size and the high complexity of the models, the risk of overfitting is severe. The authors must discuss the generalization potential of these models in this context.

Conclusion on the first two points:

This evidence of potential overfitting is a major flaw that propagates through all results and discussions in the manuscript. Before any further analysis can be considered, the authors must demonstrate that the models (especially the DDPM) are not over-parameterized for the provided dataset and that the reported performance metrics are not the result of a model that has failed to generalize.

Additional concerns

• Perfect-Model Framework and Practical Usability (Lines 154–157): The authors describe their approach as a "perfect-model super-resolution framework" where both inputs and targets originate from the same dataset (ERA5-Land). While this allows for controlled evaluation, it ignores the critical real-world challenge of "predictor mismatch" or bias between different datasets (e.g., GCM output vs. regional simulations). This design choice significantly hinders the practical usability of the proposed models, as they are not tested against actual biased low-resolution information found in climate model outputs. This limitation must be explicitly stated in the Abstract, and discussed in the Introduction, and Conclusion. The authors should clarify how these models would perform when the "trace" of mesoscale information is truly absent or biased in the low-resolution input.
• Data Normalization: The manuscript mentions a log(1+x) transformation and min-max normalization specifically for the DDPM (Line 258). It is unclear if any normalization or transformation was applied to the precipitation data for the U-NET and WGAN models. Given the heavy-tailed nature of precipitation, this is a critical detail. If no normalization was used for the other models, the authors must justify this choice or clarify the preprocessing steps for all architectures.
• Stochastic vs. Epistemic Uncertainty: The authors use 10 independently trained models to form an ensemble (Line 270), which they state reflects "epistemic uncertainty." However, the primary strength of generative models like WGAN and DDPM is their ability to produce multiple stochastic realizations from a single trained model. The authors should evaluate the generative potential/stochasticity of a single trained model rather than relying solely on an ensemble of differently trained models, as the latter is computationally expensive and less practical for end-users.
• Static Variables: Were any static variables (e.g., elevation, land cover, distance to coast) included as auxiliary inputs to the models? Orographic forcing is mentioned as a relevant driver (especially for one of the domains/regimes line 127), but looks like the models were trained without providing them with this information. Having access to the underlying topography can assist in the downscaling process: why did the authors choose not to include any static covariants in the training?

3. Specific Comments and Technical Corrections

• Hydrologic Relevance: The authors mention "hydrologic modeling" as a motivation many times in the text. Could the authors clarify the specific "hydrologic" metrics or assessments used beyond standard meteorological diagnostics?
• Target Data Selection (Lines 115–122): The authors use ERA5-Land as the target data. Please provide a brief justification for using a reanalysis product as "Ground Truth" (GT) rather than observational-based data (e.g., high-resolution radar products).
• DDPM Hyperparameters (Section 2.2.3): training details for the DDPM are missing. Specifically, please provide the optimizer used (e.g., Adam, AdamW) and the specific learning rate or learning rate schedule employed during training.
• Figure 3 Clarification: What do the columns in Figure 3 represent? Please label them clearly in the figure or the caption. The low-resolution input in the last column does not appear to match the high-resolution Ground Truth (GT). Please verify if the block-averaging or interpolation used to create the low-resolution inputs is consistent across the visualization.