acceptodds
Under review as a conference paper at ICLR 2027

MUNDO: A Multiresolution Generative Framework for Temporal Downscaling

Abstract

Dynamical systems are often observed at coarse temporal resolution through sparse states or temporal aggregates, leaving the underlying fine-scale trajectory unresolved. We present MUNDO (Multiresolution UNified DOwnscaling), a generative framework for temporal downscaling. MUNDO builds on engression, which trains a conditional sampler with a strictly proper scoring rule and generates a sample in a single network evaluation, without iterative diffusion-based sampling. MUNDO extends this to a hierarchy of resolutions by factorizing the observation operator into successive coarsenings. The resulting conditionals are then sampled in reverse, with each level generating only the degrees of freedom lost between adjacent resolutions. We evaluate MUNDO on two-scale Lorenz-96, four 2D physical systems, and multivariate ERA5 reanalysis over Europe, for both temporal interpolation and disaggregation. Across datasets and tasks, MUNDO generally outperforms flow-based baselines. Single-step MUNDO achieves strong predictive skill, while recursive MUNDO generally improves ensemble calibration, with better spread-skill agreement and predictive coverage.

Primary area:
learning on time series and dynamical systems
Keywords:
temporal downscaling, inverse problems, generative models, dynamical systems, proper scoring rules

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

This paper introduces MUNDO, a generative framework for temporal downscaling under two linear observation operators: subsampling (temporal interpolation) and block averaging (temporal disaggregation). The framework combines energy-score-based conditional generation with a null-space parameterization to enforce strict observation consistency. Its recursive variant factorizes temporal coarsening across successive resolution levels and learns the corresponding reverse conditional distributions. Empirical evaluations span the two-scale Lorenz-96 model, four simulated physical systems, and multivariate ERA5 reanalysis data. While the paper tackles a practically important problem—particularly regarding single-step generation and sampling efficiency compared to flow-based baselines—the current manuscript does not sufficiently differentiate its contributions from closely related work on hard-constrained generation and multiresolution distributional learning. Furthermore, several experimental comparisons confound the effects of recursion with variations in model capacity, training budget, conditioning information, and evaluation protocols. Consequently, these confounding factors limit the conclusions that can be drawn regarding both the framework's methodological novelty and the true benefits of its recursive formulation.

Critical strengths
  1. Unified operator-consistent formulation with strong single-step efficiency: Formulating temporal interpolation and disaggregation under a common linear observation model (y=Axy = Ax) clearly highlights the variance deficit of deterministic regression and motivates conditional trajectory sampling. By combining energy-score distributional regression with an explicit range–null space decomposition, x^=A†y+(I−A†A)Gθ(y,ε,c)\hat{x} = A^\dagger y + (I - A^\dagger A)G_\theta(y, \varepsilon, c), MUNDO guarantees exact observation consistency (Ax^=yA\hat{x} = y) without iterative likelihood-gradient updates. This is consequential to my assessment because single-step MUNDO (R=1R=1) already delivers strong accuracy and substantial sampling speedups relative to the evaluated iterative flow-matching baselines.

  2. Systematic evaluation across dynamical regimes and informative component ablations: The empirical study spans a well-chosen hierarchy of dynamical systems—from two-scale Lorenz-96 with controllable chaoticity to four 2D PDEs across varying decorrelation horizons (Nδ/τN\delta/\tau) and multivariate ERA5 reanalysis. Furthermore, the appendices transparently investigate critical design choices—including projection placement during training vs. sampling, spatial-coherence regularizers, cross-window conditioning, and end-to-end fine-tuning—while candidly acknowledging regimes where the recursive hierarchy underperforms direct single-step generation.

Critical weaknesses
  1. Insufficient positioning and differentiation against closely related literature: The manuscript does not adequately situate its architectural and methodological contributions relative to established work on hard-constrained downscaling, proper-scoring generative models, and atmospheric inverse problems:

    • Prior hard-constrained and single-pass temporal downscalers: The introduction claims that no existing method guarantees exact observation consistency without an iterative solve. However, RainDisaggGAN [1] already achieves single-pass stochastic temporal disaggregation of precipitation with exact aggregate preservation via a temporal softmax-and-scaling layer. More broadly, Harder et al. [2] formulated additive hard-constraint layers whose algebraic form is identical to the block-average residual projection used in MUNDO. Meanwhile, HourGlass [5] (cited only briefly) already performs single-pass probabilistic temporal downscaling trained with fair CRPS and auxiliary penalties.
    • Overlap with multiresolution proper-scoring frameworks: While the paper cites engression and Reverse Markov Learning [4], it omits EnScale / EnScale-t [3], which already integrates energy-score training, progressive multiresolution refinement, and autoregressive conditioning on preceding states at each resolution level. Consequently, the boundary between inherited components [2, 3, 4, 5] and genuinely new algorithmic contributions in MUNDO remains unclear, and the paper lacks discussion of how its task-specific training trades off against flexible zero-shot atmospheric inverse solvers such as WIND [6].
  2. Framework design tensions, confounded experimental protocols, and paradoxical recursive results: Several inconsistencies in framework design and experimental control undermine the paper's core claims regarding the benefits of the recursive multiresolution construction (R>1R > 1):

    • Design tensions in projection placement, error accumulation, and non-linear constraints: First, single-step MUNDO trains on unprojected outputs and only projects at inference, because applying the null-space projection inside the training loss consistently degrades CRPS (Appendix A.3, Table 6), whereas recursive MUNDO trains on projected outputs. The paper does not explain why end-to-end null-space projection hurts optimization in R=1R=1. Second, rather than simplifying learning, level-wise factorized training suffers from severe exposure bias and high-frequency artifacts (e.g., Table 5 and Table 8 at N=8N=8), requiring expensive end-to-end joint backpropagation through all RR networks with a different loss (spectral CRPS) to remain competitive. Third, linear affine projection onto ker⁡A\ker A ignores domain positivity and non-linear state coupling: for non-negative variables like ERA5 precipitation or coupled fields like density and ∣ψ∣2|\psi|^2 in Gross-Pitaevskii, subtracting a block mean in ker⁡Aavg\ker A_{\text{avg}} can induce negative values or physical inconsistencies that cannot be naively clipped without violating Ax^=yA\hat{x} = y.
    • Unexplained degradation of recursion and confounded experimental controls: Empirically, single-step MUNDO (R=1R=1) consistently beats recursive MUNDO (R>1R>1) in CRPS across nearly all interpolation settings and most AutoSim disaggregation regimes. Where recursive MUNDO does show gains, the experimental design conflates recursion with: (i) 4x model capacity and 3.7x training compute on ERA5 (4×57.4M4 \times 57.4\text{M} vs. 1×57.4M1 \times 57.4\text{M}, Table 11); (ii) asymmetric conditioning and objectives, where only R>1R>1 receives generated predecessor states (bt−1(r)b_{t-1}^{(r)}) and Stage-2 spectral-CRPS fine-tuning, and where baseline access to neighboring coarse observations (yt−1,yt+1y_{t-1}, y_{t+1}) is not clearly equalized; (iii) mismatched evaluation targets on Lorenz-96 (Appendix B.1), where causal-chain recursive disaggregation is scored on 175175/7575 contiguous chain windows (M=128M=128) while R=1R=1 and baselines are scored on 5050 separated windows (M=256M=256); and (iv) missing single-pass baselines (e.g., an HourGlass-style model or hard-constrained cGAN) using the exact same backbone and final projection ✓π\checkmark^\pi.

References:

[1] Scher, S., & Peßenteiner, S. (2021). Technical Note: Temporal disaggregation of spatial rainfall fields with generative adversarial networks. Hydrology and Earth System Sciences, 25(6), 3207–3225.

[2] Harder, P., et al. (2023). Hard-constrained deep learning for climate downscaling. Journal of Machine Learning Research, 24(365), 1–40.

[3] Schillinger, M., Samarin, M., Shen, X., Knutti, R., & Meinshausen, N. (2026). EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rules. Journal of Advances in Modeling Earth Systems, 18(8)

[4] Shen, X., Meinshausen, N., & Zhang, T. (2025). Reverse Markov learning: Multi-step generative models for complex distributions. arXiv

[5] Ingstad, M. S., et al. (2026). HourGlass: A probabilistic data-driven temporal downscaler for global and regional weather forecasting. arXiv preprint arXiv:2607.11457.

[6] Aich, M., Fürst, A., Sestak, F., Ruiz-Gonzalez, C., Boers, N., & Brandstetter, J. (2026). WIND: Weather inverse diffusion for zero-shot atmospheric modeling. In Proceedings of the 43rd International Conference on Machine Learning.

Questions for authors
  1. Beyond combining established additive/null-space constraints with reverse Markov energy-score learning , what is the core methodological innovation of MUNDO?
  2. Why does recursive MUNDO (R>1R>1) underperform single-step MUNDO (R=1R=1) in CRPS across most interpolation and AutoSim settings, and why does projecting inside the training loss degrade R=1R=1 (Table 6) while being used for R>1R>1?
  3. Do the recursive gains on ERA5 and AutoSim persist when total parameter count, training budget, spectral-CRPS fine-tuning, and generated predecessor context are strictly matched with R=1R=1?
  4. Were all baselines given the exact same neighboring coarse observations (yt±1y_{t\pm 1}) and static/calendar inputs as MUNDO, and can the Lorenz-96 disaggregation methods be evaluated on identical test windows and ensemble sizes (with post-projection ✓π\checkmark^\pi applied to baselines)?
3 comments

We thank the reviewer for the thorough review. We address novelty and positioning (Q1), projection (Q2), recursive gains and experimental controls (Q3), baseline inputs and evaluation protocol (Q4), and positivity. Due to the character limit, our response is split across two comments.

Novelty and positioning (Q1). MUNDO builds on engression and reverse Markov learning, and extends them to temporal downscaling through four contributions: (i) a null-space parameterization under a general linear observation operator (Eq. 9), which gives exact consistency with the observation in one network evaluation per draw, for interpolation and disaggregation alike (Eq. 4); (ii) a multiresolution factorization of that operator (Eq. 10) in which each level is exactly consistent with its parent, so that consistency propagates along the chain; (iii) end-to-end fine-tuning of the levels on their own generated inputs; and (iv) a spatial-coherence term complementing the energy score. The revised introduction lists them. What we propose is this framework as a whole, kept unchanged across three backbones and six systems (Table 1).

Our novelty claim concerns methods that generate samples satisfying the observation exactly, under both temporal observation operators, without an iterative solve. The closest works are discussed in Section 2 and Appendix A.2, and compared in Table 5:

  • RainDisaggGAN enforces exact aggregates via softmax rescaling. It does so for a single operator (aggregation) and a single non-negative variable. Our ERA5 experiments cover seven fields and both operators.
  • Harder et al. (2023) address spatial downscaling. Their temporal model is deterministic, and its constraints act on spatial means. For block averaging, their additive constraint layer coincides with our projection, and we now cite it. Our null-space parameterization only requires AA to be linear with full row rank. The same construction therefore applies to subsampling and to the per-level operators of our hierarchy, without operator-specific design.
  • EnScale shares engression and the coarse-to-fine approach, but addresses spatial downscaling. Its stages condition on block-averaged targets without imposing them. MUNDO's hierarchy factorizes the observation operator, so each level is exactly consistent with its parent, and spatial coherence is encouraged through an additional score term (Appendix B.2).
  • HourGlass is concurrent and the closest in training: a single-pass generator trained with a CRPS-type score. It handles the full atmospheric state (about 100 fields), but generates the five hourly states between two six-hourly states, for interpolation only, and penalizes deviations in window statistics instead of imposing them. Our ERA5 windows span 24 hours under both operators, and our sweeps extend to windows of 128 steps and 16.6 decorrelation times (Appendix D.1).
  • Inverse solvers. Unlike DDNM and FlowDPS, MUNDO requires neither iterative sampling nor consistency gradient steps. FlowDPS represents methods using an unconditional prior trained on the target data, with the observation imposed at sampling time. WIND is a zero-shot solver built on a pretrained foundation model. Its released checkpoint differs from our ERA5 setting (0.25∘0.25^\circ, hourly, seven fields) in resolution (1.5∘1.5^\circ), temporal frequency (6-hourly), and number of variables (70). Retraining its prior is beyond our computational budget (Appendix A.2).

Projection (Q2). Recursive levels are parameterized in the null space by construction. Each network predicts only the free values at its level, and children are assembled to match their parent exactly (i.e., the network adds a residual that the coarsening maps to zero, e.g., residuals that sum to zero around the parent mean value under disaggregation). There is therefore no separate projection step, and projected samples are used directly in the training objective.

For single-step MUNDO, projecting during training or only at sampling yields the same population optimum under finite first moments and realizability (Appendix B.3). With finite-capacity networks, the objective does affect the learned distribution. On AutoSim disaggregation at N=32N=32, projecting inside the training loss increases CRPS on all four systems, with mixed effects on calibration and little change in temporal structure (Table 7). We therefore project only at sampling time for single-step MUNDO.

We address here the remaining points: recursive gains and experimental controls (Q3), baseline inputs and evaluation protocol (Q4), and positivity.

Recursive gains and experimental controls (Q3). Single-step MUNDO often achieves lower CRPS than recursive MUNDO on interpolation and AutoSim disaggregation. The hierarchy mainly improves calibration and temporal structure, which CRPS alone does not measure. Its gain is largest on ERA5, the benchmark with the most coupled variables (seven fields; Table 3).

  • Capacity. On Lorenz-96 and AutoSim, both MUNDO variants and both flow-based baselines share the same backbone and are matched in parameter count (Appendix C.4, Table 13).
    • On ERA5, the recursive model has about 4$\times$ more parameters (230M across four levels). Matching it with a single network would require 46-61 GiB peak VRAM at batch size 2 on an 80-GB H100, and exceed memory at batch size 4, the batch size of all our ERA5 runs. We therefore did not run these variants, and the ERA5 comparison does not isolate recursion at matched capacity.
    • The hierarchy does offer a memory advantage. Each level generates only the degrees of freedom between two successive resolutions and is trained independently, so only one network has to fit in memory at a time. Total capacity can thus grow at a fixed memory budget, at the cost of RR evaluations per sample.
    • At matched capacity (57.4M parameters), single-step MUNDO already outperforms the flow-based baselines on ERA5. Its CRPS is 21% lower than FlowDPS under disaggregation and 33% lower than flow matching under interpolation (Table 3).
  • Fine-tuning, loss and predecessor context. End-to-end fine-tuning is optional and used only when validation diagnostics indicate error propagation across levels (Section 3.5).
    • The ERA5 recursive model uses neither this fine-tuning nor generated predecessor context (Table 13), so neither explains its ERA5 gains.
    • On AutoSim, replacing true with generated predecessor context changes CRPS by about 2% or less, except for advection-diffusion at N=8N=8 (Table 10).
    • The method does not require a specific loss either. On Gross-Pitaevskii disaggregation at N=24N=24, CRPS stays within 1.4% across the three loss configurations (0.0788 to 0.0799; Table 6).
  • Single-pass constrained control. Single-step MUNDO provides this control. It uses the same backbone and final projection, and differs from a HourGlass-style model in its training objective and in imposing consistency exactly. We did not run a hard-constrained cGAN: the softmax constraint of RainDisaggGAN applies only to the aggregate of a non-negative variable, so it covers neither interpolation nor signed fields such as the wind components.

Baseline inputs and evaluation protocol (Q4). The reviewer asks whether the baselines receive the same inputs as MUNDO, and whether the Lorenz-96 methods can be evaluated on identical windows with projected baselines. Yes on both counts (Section 4.1, Appendix C.3):

  • All learned generative methods receive the hour, day of year, and four static fields. SimVP, following its original design, receives only the observation-derived input video.
  • For disaggregation, MUNDO and both flow-based baselines also receive the aggregates of adjacent windows.
  • On AutoSim and ERA5, learned baselines were already projected onto the constraint set after sampling. We now apply the same projection on Lorenz-96.
  • On Lorenz-96, all methods are now evaluated on the same 25 chains (Appendix C.1).

Positivity. The reviewer is right that linear aggregate constraints do not guarantee non-negativity, and this affects all methods. The share of negative precipitation pixel-hours is 17% for Denton-Cholette, 20% for flow matching, 24% for single-step MUNDO, and 28% for recursive MUNDO.

We address this by projecting precipitation onto the intersection of the aggregate constraint and the non-negative orthant, x:Ax=y,x≥0\\{x: Ax=y,\\ x \geq 0\\} (Appendix B.3). This projection has a closed-form blockwise solution up to a scalar threshold found by bisection, with negligible sampling cost. Applied to all methods, it removes negative values while preserving exact daily means. For MUNDO, it also improves precipitation scores under disaggregation: CRPS drops by 5% for the recursive model and 4% for the single-step model, and the finest spectral band ratio and t1t_1 of the recursive model move close to one. Its spread-skill ratio decreases, since part of the spread came from negative values.

All ERA5 precipitation results in the revision use this projection, and Table 16 (Appendix D.3.1) compares every method before and after. Alternatively, a log⁡(x+ε)\log(x+\varepsilon) parameterization could enforce positivity directly and is left for future work. For coupled fields such as density and ψ\psi in Gross-Pitaevskii, we claim consistency with the observation operator, not preservation of physical invariants.

Thank you for the thorough response. I will reconsider my score.

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

The paper presents MUNDO, a probabilistic temporal downscaling framework that unifies interpolation and disaggregation under linear observation operators. It combines engression-based conditional generation, energy score training, and null space projection to enforce consistency with coarse observations. Building on reverse Markov learning, it introduces a temporal multiresolution implementation that generates fine-scale details progressively while preserving consistency with each coarser parent. The framework is evaluated on two-scale Lorenz–96, four AutoSim physical systems, and ERA5 reanalysis data, with comparisons against generative and deterministic baselines and additional component ablations. The results demonstrate competitive single-step performance and calibration benefits from recursive sampling in several settings, while the predictive gains from the multiresolution structure vary across tasks and systems.

Critical strengths
  1. Broad empirical evaluation with promising results. The paper evaluates the framework on controlled dynamical systems, four two-dimensional physical systems, and ERA5 reanalysis data, covering both interpolation and disaggregation. Additional experiments vary window lengths and predictability regimes and examine individual components. The reported results demonstrate competitive probabilistic performance in several settings, with calibration benefits from recursive sampling.
  2. Explicit observation consistency and efficient sampling. The framework handles interpolation and disaggregation within a common formulation and uses null-space projection to enforce the linear observation constraints exactly. In the recursive version, this consistency is maintained at each resolution level. Sampling requires only one network evaluation per level, providing a practical efficiency advantage over the evaluated iterative flow-based baselines
Critical weaknesses
  1. The figures do not clearly communicate differences from the baselines. In several figures (e.g., Figures 3, 8, 9, 12, and 13), the results of MUNDO and competing methods appear visually similar, making it difficult to identify the proposed method’s advantages. Dense layouts and small panels further limit comparison.
  2. The empirical contribution of multiresolution sampling remains insufficiently established. Progressive sampling across temporal resolutions is presented as a main contribution, but its benefits over the single-step version are mixed across settings and evaluation metrics. Improvements in some metrics can accompany deterioration in others, leaving the conditions under which the additional hierarchy is beneficial unclear. The paper provides limited analysis of how performance depends on the number of levels (R), or guidance for selecting it.
Questions for authors

Impact of temporal multiresolution factorization and guidance for its selection. How does performance across the evaluation metrics vary with the number of resolution levels when the task and window length are held fixed? Could the authors clarify when additional levels are beneficial and provide empirical evidence supporting these choices?

Clearer visualization of improvements over the baselines. Could the authors use more informative visualizations to clarify where and to what extent MUNDO improves over the baselines?

2 comments

We thank the reviewer for the careful reading and positive assessment. We address the two main questions raised: how to choose the number of levels RR, and how multiresolution sampling improves over single-step sampling. We add an experiment on the role of RR and a figure that shows all metrics at once.

Number of levels RR. In our construction, RR is determined by the factorization of the coarsening ratio kk. We use its finest integer factorization, with one level per factor (e.g., k=24=2⋅2⋅2⋅3k=24=2 \cdot 2 \cdot 2 \cdot 3 on ERA5). Fewer levels correspond to grouping these factors into larger coarsening steps, which makes the order of the factors a design choice (e.g., 4 then 8, or 8 then 4).

Therefore, we add a controlled sweep over RR on Navier-Stokes disaggregation at N=32N=32 (Appendix B.4, Table 8, Figure 7). It covers 11 chains from R=1R=1 to R=5R=5, with parameter counts within about 10% of the R=5R=5 model. More levels improve calibration at the expense of CRPS: from R=1R=1 to R=5R=5, the coverage error falls from 0.095 to 0.020 while CRPS rises by 23%. Additional levels are therefore beneficial when calibration matters more than CRPS. We use the finest factorization by default, since it merges no factors and so avoids the ordering ambiguity.

When recursion helps. We agree that the benefits of multiresolution sampling are mixed across metrics: single-step MUNDO often has the lower CRPS, in particular under interpolation. CRPS is marginal in time and space, and does not measure calibration or temporal structure. The gain of recursion is mainly in calibration and temporal structure. For calibration, the sweep above shows the gain at matched capacity, and the spread-skill ratio of recursive MUNDO is closer to one in 37 of the 44 AutoSim settings (Appendix D.4).

For temporal structure, recursion keeps the lag-1 variogram ratio t1t_1 close to one where single-step MUNDO drifts: on ERA5 (1.01 against 1.19 under disaggregation, 1.04 against 1.27 under interpolation; Table 3) and on Lorenz-96 as predictability decreases (Appendix D.1). On AutoSim, both variants already keep t1t_1 close to one (Table 4). Under disaggregation, all learned methods receive the aggregates of adjacent windows. With a hierarchy, each level can also condition on the last state generated for the preceding window. This helps keep consecutive windows continuous (Figure 3, Appendix B.5).

Where predictability is lowest, recursion also lowers CRPS under disaggregation: by 19-39% on chaotic advection-diffusion at N≥16N \geq 16, and on Lorenz-96 in all settings but one (Appendix D.1).

The gap is clearest on ERA5 (Table 3, Figure 2), which couples seven heterogeneous fields, where Lorenz-96 and the AutoSim systems have one to three. There, recursive MUNDO is better than single-step MUNDO on every metric under both operators (e.g., a coverage error of 0.147 against 0.250 under disaggregation). Our hypothesis is that recursion helps most when several coupled variables and time scales are generated jointly. The recursive ERA5 model also has four times the parameters of the single-step one (Appendix C.4), so this comparison alone does not separate the two effects.

Figures. Section 4.2 now opens with radar charts (Figure 2) that display all metrics jointly for every benchmark, with the best classical baseline dashed. The scores remain in Tables 2-4. For the reconstruction results (now Figures 5, 10 and 11), we reduced the number of displayed time steps. We moved the Lorenz-96 single-variable figure to the main text (Figure 3): it compares the ensembles of each method at one Lorenz-96 variable against the reference and the observations. Figures 14 and 15 now show spectra as ratios to the reference spectrum, with one as the target.

Thank you for the detailed response and clarification. I reviewed the revision and the rebuttal, and I think they have solved my main concerns on how to choose the number of levels and the visualization of the figures. I thus maintain my score of week acceptance but would like to increase my confidence to 3.

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

This manuscript proposes a generative method for temporal downscaling. This is a classic problem, with proposed solutions from classic to recent learned method. The paper is written reasonably well.

Critical strengths

The paper proposes a new method to a classic problem. The paper is written reasonably well.

Critical weaknesses

This is a classic problem. Existing methods 1970s, Denton-Cholette, to more modern learned method. The manuscript presented the method well. However, it is also important to compare with the existing baseline. The paper should provide sufficient details to the novelty of the proposed method, comparing to the existing methods.

Questions for authors

Without uploading this manuscript to any LLM. But ChatGPT provided the following list, after being asked about recent learned temporal downscaler, which can be used for scientific machine learning applications:

  • RIFE 2022, Learned intermediate optical flow; fast interpolation
  • FILM 2022, Multiscale feature-based interpolation for large motion
  • IFRNet 2022, Intermediate feature refinement with learned flow
  • FLAVR 2021, 3D convolutional interpolation without explicit optical flow
  • AMT 2023, All-pairs correlation and multiscale motion refinement

Given the short time to complete the review, I can't judge whether the related work section contains the state-of-the-art prior methods, and also for the baseline methods. I believe the author(s) had the same challenge. But given the long history of the problem. The author(s) should use more space for related work discussion, and justify the choices of baseline methods. If space limit is a concern, a more comprehensive related work discussion can be included in the appendix.

3 comments

We thank the reviewer for the review and for the suggested references. We read its two main points as follows: classical baselines alone are not a sufficient comparison, and the novelty relative to existing approaches is unclear. This comment covers what the paper proposes and how the approach compares to the baselines. A second comment covers the suggested video methods and novelty.

What the paper proposes. MUNDO is a generative framework for temporal downscaling. We evaluate it against other approaches under a controlled protocol, across settings of increasing complexity. In each setting, we take a backbone suited to the system (a circular CNN on Lorenz-96, NATTEN on AutoSim, ArchesWeather on ERA5; Table 1), and keep the framework unchanged. This choice is consistent with recent evidence from AI weather forecasting that training methodology, loss design and data diversity matter at least as much as the choice of architecture (Garg et al., 2026, arXiv:2604.01215). We do not test that claim with a backbone ablation. We hold the backbone fixed across methods instead, as described below.

Baselines. The submission already compares the framework with one representative of each existing family (Table 5, Appendix A):

  • Classical. Denton-Cholette and CDF-t for disaggregation, linear and cubic-spline for interpolation. We keep them because statistical methods of this kind are still developed and compared against in recent work (Appendix A.1). They are evaluated on every benchmark (Tables 2-4), and Figure 2 draws the best one in each panel.
  • Conditional generative. Conditional flow matching (Lipman et al., 2023) is the flow-based counterpart of the diffusion models used for downscaling.
  • Inverse-problem solver. FlowDPS (Kim et al., 2025) imposes the observation on an unconditional prior at sampling time. It stands for the family of DDNM (Wang et al., 2023), Schmidt et al. (2025) and WIND (Aich et al., 2026).
  • Deterministic. SimVP (Gao et al., 2022) is a video prediction model. We keep its original design, so it shows how well a strong deterministic architecture performs on the same task.

MUNDO, flow matching and FlowDPS use the same backbone and are matched in parameter count on Lorenz-96 and AutoSim. On ERA5, single-step MUNDO and both flow-based baselines use the same 57.4M-parameter network (Appendix C.4, Table 13). Differences between these three methods therefore come from how the model is trained and sampled. At matched capacity, single-step MUNDO has a lower CRPS than both flow-based baselines on the four AutoSim systems under both operators (Table 4), and on ERA5 by 21% under disaggregation and 33% under interpolation (Table 3). Flow matching is ahead on Lorenz-96 disaggregation (Table 2) and on some advection-diffusion settings, where its ensembles are underdispersed (Appendix D.4).

Scope of the evaluation. The same framework is evaluated on states from 36 variables to 149×281×7149 \times 281 \times 7, and on windows from 8 to 128 steps that span up to 16.6 decorrelation times (Table 15). These sweeps probe how far it can be used in practice, from six-hourly interpolation to monthly-mean disaggregation. As predictability decreases, CRPS rises for every method, but MUNDO draws still satisfy the observation exactly and keep a realistic short-lag structure (Appendix D.1, D.4).

We address here the remaining points: the suggested video interpolation methods, and novelty.

Suggested video interpolation methods. RIFE, FILM, IFRNet, FLAVR, and AMT were developed for RGB footage, where motion is the displacement of objects in a scene. Most rely on optical flow and on assumptions such as brightness constancy and explicit occlusion handling. These assumptions do not transfer directly to geophysical fields. STint, TemDeep and ModAFNO are adaptations of frame interpolation to geospatial and climate data. We discuss them in Section 2, and Appendix A.2 covers video frame interpolation with and without optical flow. Our baseline from this literature is SimVP.

These methods share two limits in our setting: they are deterministic, and they address interpolation only. As discussed in Section 3.2, the problem is ill-posed: an observation is compatible with an affine subspace of fine-scale trajectories. A deterministic model returns one element of this set and, trained with a squared loss, converges to the conditional mean, which underestimates the variance (Eq. 7).

Novelty. MUNDO builds on engression and reverse Markov learning, and extends them to temporal downscaling through four contributions: (i) a null-space parameterization under a general linear observation operator (Eq. 9), which gives exact consistency with the observation in one network evaluation per draw, for interpolation and disaggregation alike (Eq. 4); (ii) a multiresolution factorization of that operator (Eq. 10) in which each level is exactly consistent with its parent, so that consistency propagates along the chain; (iii) end-to-end fine-tuning of the levels on their own generated inputs; (iv) a spatial-coherence term that the energy score alone lacks. The revised introduction lists these four points.

Taken together, they form the framework described in our first comment. Table 5 (Appendix A) compares MUNDO with 19 existing methods on five criteria. Among them, only MUNDO applies to both operators, returns samples of the conditional law, satisfies the observation exactly by construction, and needs no iterative solver. DDNM meets the first three but requires an iterative solve, and RainDisaggGAN is exact in one pass but restricted to the disaggregation of a non-negative variable.

Thanks for the detailed comments. I will read them through and revise my review accordingly.

Comments on the paper

1 comment

We sincerely thank the three reviewers for the time and care they put into evaluating our work, and for the discussion that followed our responses. Their comments led to concrete changes, and we believe the paper is clearer and better supported as a result. We also thank the area chairs and program chairs for running the review process.

The revision includes the following changes:

  • Contributions and positioning (reviewers hjMm, mSmW). The introduction now states our four contributions more explicitly. We rewrote the related work (Section 2) and added an extended version in Appendix A, which justifies our choice of baselines and positions MUNDO against 19 existing methods on five criteria (Table 5).
  • Number of levels RR and when recursion helps (reviewers UhAU, mSmW). To better explain the number of resolution levels, we added a controlled sweep over 11 chains from R=1R=1 to R=5R=5 at matched capacity on Navier-Stokes disaggregation (Appendix B.4, Table 8, Figure 7). More levels improve calibration at the expense of CRPS, consistent with the rest of our results, where the gain from recursion grows with the complexity of the system.
  • Evaluation protocol (reviewer mSmW). We made two changes on Lorenz-96: all methods are now evaluated on the same 25 chains, and learned baselines are projected onto the constraint set, following the protocol used on AutoSim and ERA5.
  • Positivity (reviewer mSmW). We now handle non-negative variables such as precipitation in ERA5 with a projection that removes negative values while preserving the aggregates exactly (Appendix B.3). Table 16 reports the precipitation scores of every method with and without this projection.
  • Figures (reviewer UhAU). We added radar charts that compare the methods on the main metrics jointly, for every benchmark (Figure 2). We moved the Lorenz-96 single-variable figure to the main text (Figure 3): it compares the ensembles of each method at one variable against the reference trajectory and the observations. Finally, we revised two sets of figures. The reconstruction figures display fewer time steps (Figures 4, 5, 10 and 11), and the spectra are plotted as ratios to the reference spectrum (Figures 14 and 15).

We thank the reviewers again for a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.