acceptodds
Withdrawn from ICLR 2027

Conditioning Without Guidance: When the Observation Is Already a Diffusion Marginal

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Conditioning a diffusion model usually approximates an unavailable likelihood score at inference. We identify a restricted case where this is unnecessary: when an observation has the law of a reachable forward marginal, conditioning reduces to restarting the reverse process there. With an exact score and exact marginal match, restart samples the Bayes posterior. We instantiate this idea with scale-time diffusion, which destroys Fourier modes in scale order. A two-segment schedule removes the unobserved band and then matches sensor noise. Noiseless observations remain fixed by the reverse dynamics; noisy ones are denoised before they freeze. Across two chaotic dynamical systems involving turbulent flows, the method substantially reduces reconstruction error relative to learned baselines in the first three missing octaves while retaining fine-scale power through repeated assimilation. A progressive clock uses a finer unseen observation with an exactness–utility tradeoff. The theorem covers matched projection sensors. The production endpoint has a climatology-averaged joint-law total-variation bound of ; sparse sensing and useful off-design entry remain outside exact matching.

Primary area:
generative models
Keywords:
diffusion models, score-based generative models, inverse problems, data assimilation, super-resolution, scientific machine learning, uncertainty quantification, neural operators

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_jFqROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a data assimilation method that regards the observation as a diffusion state in the forward process. By doing so, it can simply reverse the diffusion from that observation without using DPS style gradient or guidance. The experiments in turbulence and surface quasi-geostrophic reconstruction task show it outperform other baseline including EnKF, IR Diffusion, FNO and other baselines.

Critical strengths

The method proposed in this paper provides a conditioned diffusion method that can using observation directly in data assimilation or super resolution task in inference level rather than previous work such as Score-based Data Assimilation. Which could be a new way for diffusion based data assimilation task.

Critical weaknesses

Firstly, some parts in this paper are not well organized. In particular, Figure 1, which is the main figure of this paper, have many issues. This figure contains patches with different font styles and different resolutions. Some parts are blurry and visually inconsistent, which makes the figure look like it was not carefully prepared. The writing has similar presentation issues and make the paper difficult to follow.

Secondly, there have already been several works on conditioned diffusion for PDE solving and data assimilation, for example, On Conditional Diffusion Models for PDE Simulations [1]. Therefore, the actual novelty of this work should be discussed more clearly in relation to these existing approaches.

Lastly, the baselines and experimental evaluations are relatively simple and limited the author should using more recent data assimilation methods such as SDA [2] to compare with it own method. And test it on more experiment with different physical domain to avoid conclusions that may be specific to the selected test cases

[1] Shysheya, Aliaksandra, et al. "On conditional diffusion models for PDE simulations." Advances in Neural Information Processing Systems 37 (2024): 23246-23300.

[2] Rozet, François, and Gilles Louppe. "Score-based data assimilation." Advances in Neural Information Processing Systems 36 (2023): 40521-40541.

Questions for authors

Could the authors clarify how Figure 1 was prepared?

Could the authors explain why recent score-based data assimilation methods, such as SDA, are not included as baselines?

Could the authors compare the proposed method with other conditional diffusion methods for data assimilation?

Reviewer_wDQoOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a scale-time diffusion method for super-resolved data assimilation. By matching the forward diffusion process to the observation model, the method performs posterior sampling without additional likelihood guidance. It uses low-resolution spectral observations and FNO forecasts to reconstruct high-resolution states. Experiments on 2D turbulence and SQG show improvements in fine-scale reconstruction and stable performance in cycled assimilation.

Critical strengths
  1. Effective approach. The method combines super-resolution with data assimilation in a principled way. It achieves better reconstruction accuracy than the learned baselines, particularly at unresolved scales, while preserving fine-scale spectral power.

  2. Extensive experiments and stable cycling. The paper includes comparisons with several baselines, ablation studies, observation-noise experiments, and probabilistic evaluation. The method is also evaluated in repeated forecast–analysis cycles rather than only single-step reconstruction. The results show stable cycling and lower long-term reconstruction errors than the learned baselines.

Critical weaknesses
  1. Limited applicability to practical observations. The method relies on spectral observations that match the predefined diffusion process. It cannot directly handle sparse spatial observations or general observation operators, which are common in practical data assimilation. The flexibility is also limited when the observed frequency range changes. It is unclear what real-world applications would satisfy these assumptions.

  2. Limited advantages over traditional DA methods. The method does not clearly outperform EnKF. The overall errors are comparable to those of the perfect-dynamics EnKF, while EnKF performs better in the first unresolved frequency band. Although the forecast models differ, the practical benefit of using this more complex diffusion-based approach is not fully established.

Questions for authors
  1. What practical applications do the authors envision for this method? Could it be extended to sparse spatial observations or more general observation operators?

  2. The paper reports CRPS and spread/skill, but does not clearly compare the RMSE of individual diffusion samples and their ensemble mean. Could the authors provide this comparison to assess how much ensemble averaging improves reconstruction accuracy? How would the generated ensemble perform in subsequent probabilistic forecasts?

  3. The paper examines different observation-noise levels, but how does the method compare with IR Cascade and EnKF under the same noisy observation settings? Does its advantage persist as observation noise increases?

  4. Could the authors provide stronger baseline comparisons? The deterministic SR baseline does not use forecast information, unlike the proposed method. The EnKF settings are not systematically tuned, and IR Cascade is not retuned for SQG. Could the authors evaluate whether the improvements persist under more comparable baseline settings?

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

The paper identifies conditions under which diffusion-based conditioning can be performed without an inference-time likelihood guidance term. When an observation-conditioned entry state matches the forward corruption process, exact reverse-time sampling with an exact score produces the posterior conditioned on the observation and forecast context. The authors instantiate this principle through scale-time diffusion, which progressively destroys unobserved Fourier modes and separately introduces noise into observed modes to match the sensor noise. The method is evaluated on super-resolution data assimilation for two turbulent systems, including repeated assimilation cycles, spectral reconstruction, calibration diagnostics, and comparisons with learned reconstruction methods and EnKF variants.

Critical strengths
  1. A principled construction that removes a separate conditioning approximation. The paper clearly identifies a restricted setting in which observation conditioning can be incorporated into the reverse process’s initialization. Under the stated matching assumptions, this avoids approximating the generally intractable intermediate-time likelihood score and eliminates an additional guidance-weight tuning parameter. I find this explicit connection between the observation process and the generative forward process conceptually valuable. The paper also distinguishes the exact result from finite-endpoint and learned-score approximations.

  2. Meaningful validation and potential relevance beyond the demonstrated application. The experiments examine repeated assimilation, scale-resolved reconstruction, and uncertainty calibration rather than relying solely on aggregate reconstruction error or stability. The construction may also be relevant to frequency-domain inverse problems beyond fluid dynamics, such as audio bandwidth extension and suitably matched MRI acquisition settings. These applications are not demonstrated here and would require adaptation, but they suggest that the underlying principle could have broader utility than the specific assimilation experiments.

Critical weaknesses
  1. The applicability conditions are restrictive. Exact restart requires an observation operator and noise model that match a reachable state of the trained forward process. The demonstrated spectral projection satisfies this structure, whereas sparse point observations, aliased averaging, and general nonlinear observations do not. Consequently, the current method is better supported as a specialized super-resolution assimilation approach than as a general-purpose data assimilation framework. Extending the principle to practical acquisition processes remains an important unresolved issue.

  2. The empirical advantage is qualified by baseline and statistical limitations. EnKF covariance settings are not validation-tuned, and the SQG cascade comparison uses augmentation settings selected on the other system. Moreover, the 40 paired starts are drawn from only five trajectories without accounting for clustering, while the long-horizon study uses three starts. These limitations weaken conclusions about the breadth and statistical reliability of the advantage. The method is also slower than the refinement cascade, and the noiseless guidance diagnostic shows that an appropriately tuned DPS update can match its missing-band reconstruction error.

Questions for authors
  1. Benefit relative to guidance. Beyond the noiseless diagnostic, how does restart compare with a well-tuned guidance-based sampler under noisy observations and repeated assimilation at comparable computational budgets? This would help establish whether removing the guidance approximation yields measurable accuracy or calibration benefits in addition to eliminating weight tuning.

  2. Robustness of the empirical comparisons. Do the main conclusions persist with validation-tuned EnKF settings, system-specific tuning of the SQG cascade, and uncertainty estimates that account for dependence among starts from the same trajectory?

  3. Sensitivity to observation mismatch. How rapidly do reconstruction quality and calibration deteriorate when the observation departs slightly from an ideal spectral projection, for example through a smooth low-pass response? Evidence on this sensitivity would materially clarify whether the proposed principle remains useful for realistic acquisition processes.

Reviewer_e6eoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors have proposed special class of observations: spectral observation, which returns the field that retains only the low wavenumber fields. By using the wavenumber-dependent noising schedule (scale-time diffusion), the posterior for the observation operator can be identified as the marginal of an intermediate step of the diffusion process. Therefore, the posterior can be exactly computed by doing the reverse process.

Critical strengths
  • Conditioning methods without approximations such as guidance is of high importance.
  • The scale-dependent clock and the low wavenumber-picking observation operator are matching, justifying the mechanism.
Critical weaknesses
  • The novelty seems not very significant. The scale-dependent clock seems to have several precedents (e.g., Hoogeboom & Salimans, Rissanen et al. 2023). A possible novelty is that the authors pointing out that the posterior for the spectral observation operator in Definition 1 satisfies to be the marginal at the intermediate step of the diffusion process with the scale-dependent clock, but then that seems to be of minor novelty.
  • The paper writing is not very clear. For example, the third paragraph in the introduction was very hard to digest: it gives some concrete numbers like “conditioning value to 2.8×10−82.8\times 10^{-8}” and “full sampler-state residual remains 8.0×10−38.0\times 10^{-3}” without introducing how they have obtained these values. What are these values? What do the authors want to say in this paragraph?
  • Abstract is split to three paragraphs despite the ICLR template specifies “The abstract must be limited to one paragraph.”, which is a minor problem but cast me a doubt that the authors have never proofread the manuscript by themselves.
Questions for authors
  • Can you clearly state and justify the novelty of the method?

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

Related papers

Discussion (0)

Sign in to comment.