acceptodds
Under review as a conference paper at ICLR 2027

When Should Measurements Speak? Directional Guidance for Diffusion Inverse Problems

Abstract

Diffusion inverse solvers combine learned priors with measurement guidance, yet most rules use one scalar per reverse step or fixed measurement geometry. We introduce trajectory-conditioned observability (TCO), the spectrum of a noise-whitened measurement Jacobian restricted to the diffusion prior's current uncertainty subspace. In the exact local linear model, TCO yields the unique linear minimum-MSE correction and, under Gaussianity, the posterior mean and shrinkage. We prove an exact lower bound: whenever two active TCO singular values differ, every scalar measurement-gradient rule incurs strictly positive excess Bayes risk. For nonlinear operators and estimated geometry, a perturbation theorem separates curvature, subspace, residual, and trust-region errors. These results motivate DynaObs, a training-free plug-in that refreshes local geometry and applies mode-wise guidance. Across linear and nonlinear inverse problems, trajectory spectra predict failures better than endpoint diagnostics, and DynaObs improves matched-compute reconstruction while geometry-mismatched controls remove the gain.

Primary area:
generative models
Keywords:
Diffusion models, Inverse problems, Measurement guidance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QKkbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces trajectory-conditioned observability and uses it to develop DynaObs, a training-free method that adapts measurement guidance to the diffusion prior’s evolving uncertainty directions.

Critical strengths

Its exact excess-risk analysis establishes the limitations of scalar guidance within a local linear model, providing a clear theoretical motivation for adapting guidance separately along different uncertainty directions.

Critical weaknesses
  • Computational concern and lack of comparison. The repeated denoiser JVPs required to refresh the prior geometry may introduce substantial overhead, even for linear measurements, where only the measurement-side Jacobian products simplify. Moreover, Table 2 omits DDRM and DDNM/DDNM+, leaving the benefit of the evolving geometry over established spectral or range–null-space corrections insufficiently demonstrated. The authors should clarify these methodological differences and compare reconstruction quality against wall-clock time using the same pretrained model, operators, and noise levels, explicitly including geometry-estimation costs rather than relying on denoiser NFE alone.
  • Incomplete statistical reporting. Sections 6.1 and K.11 specify paired bootstrap confidence intervals, but the reconstruction and negative-control tables report only point estimates, making the reliability of the reported improvements difficult to assess.
  • Ambiguous evaluation definitions. Appendix K.5 does not specify whether Table 8 uses the largest or averaged principal angle, spectral or Frobenius projector error, or matched versus sorted spectra, limiting the interpretability and reproducibility of the geometry-estimation results.

References

  1. Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising Diffusion Restoration Models. NeurIPS, 2022. Paper
  2. Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. ICLR, 2023. Paper — introduces both DDNM and DDNM+.
Questions for authors

See weakness

Reviewer_iFm8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how measurement guidance should be distributed across directions in diffusion-based inverse problem solvers. It defines trajectory-conditioned observability, the spectrum of the noise-whitened measurement Jacobian restricted to a low-rank approximation of the denoiser's posterior covariance, derives the corresponding per-mode LMMSE correction, and shows that a single scalar guidance step incurs excess risk when this spectrum is anisotropic. The proposed plug-in, DynaObs, periodically re-estimates this geometry and applies the per-mode correction. Experiments on FFHQ and ImageNet report gains over several baselines on linear and nonlinear tasks, along with failure-prediction and negative-control studies.

Critical strengths

The motivating question (whether a single scalar guidance weight can balance prior and measurement across directions that carry different amounts of information) is relevant, and the per-direction view gives an interpretable uncertainty map. The negative controls are a good way to attribute gains to the proposed mechanism. However, these do not outweigh the weaknesses below.

Critical weaknesses

W1. Poorly written, with key details and evidence missing.

The core idea is simple (see W2) but is hard to extract from the paper, which introduces many new names for closely related quantities and relies on a long, repetitive appendix. Essential details are also missing. For example, the base sampler that DynaObs plugs into is never specified, and several quantities used in the tables and figures are not clearly defined (e.g., γt\gamma_t).

The empirical evidence is also not thorough:

  • There are no qualitative reconstructions in the paper, so it is impossible to judge whether the metric gains correspond to visibly better images.
  • While the paper reports spectral and diagnostic analyses, it never shows how the correction changes reconstructions (e.g., by visualizing the corrected directions or comparing outputs with and without the correction).
  • No uncertainty estimates are reported, although a bootstrap protocol is described, and several described experiments have no reported results.
  • "Matched compute" is claimed, but DynaObs is the slowest method in the main table and no compute-matched comparison is shown.

W2. Limited novelty, and limited theoretical support.

  • The DynaObs correction is the standard Gaussian conditional-mean (Kalman-type) update with a low-rank approximation of the denoiser's posterior covariance. Closely related covariance-aware guidance methods, such as ΠGDM [1], Peng et al. [2], TMPD [3] and MMPS [4], are neither cited nor compared, and the paper does not argue what its low-rank truncation gains over them.
  • The theory is per-step, while DynaObs reuses the same measurement at every step, and nothing connects the local results to the behavior of the full sampler.

References

[1] Song et al., "Pseudoinverse-Guided Diffusion Models for Inverse Problems," ICLR 2023.

[2] Peng et al., "Improving Diffusion Models for Inverse Problems Using Optimal Posterior Covariance," ICML 2024.

[3] Boys et al., "Tweedie Moment Projected Diffusions for Inverse Problems," TMLR 2024.

[4] Rozet et al., "Learning Diffusion Priors from Observations by Expectation Maximization," NeurIPS 2024.

Questions for authors
  1. Which base sampler is used, and does it apply its own measurement guidance in addition to the DynaObs correction? How is the scale of the estimated covariance calibrated?
  2. How were the baseline numbers obtained (re-run or taken from the original papers, and under which settings)?
  3. How does DynaObs compare with covariance-aware methods such as ΠGDM, TMPD or MMPS at matched compute?
  4. Please add qualitative reconstructions, including failure cases.
Reviewer_48ttOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces trajectory-conditioned observability (TCO) to characterize directional measurement information during diffusion-based inverse solving. It derives a locally optimal linear correction, establishes an excess-risk lower bound for scalar guidance, and proposes DynaObs as a training-free implementation. Experiments on FFHQ-256 and ImageNet-256 demonstrate improvements across linear and nonlinear inverse problems.

Critical strengths
  1. TCO provides a useful perspective on the interaction between measurement geometry and diffusion-prior uncertainty. The scalar-guidance lower bound offers a clear theoretical motivation for directional correction.
  2. The method achieves competitive results on multiple inverse problems. The mechanism ablations and geometry-mismatched controls provide useful evidence for the proposed directional guidance.
Critical weaknesses
  1. The high-SNR asymptotic expressions in Eqs. (90) and (152) appear inconsistent with the exact excess-risk formula. This affects the claimed interpretation of scalar guidance in the strong-measurement regime.
  2. The theoretical optimality relies on accurate local prior uncertainty and independence assumptions. It remains unclear whether the practical denoiser-based geometry estimates adequately satisfy these assumptions along measurement-guided trajectories.
Questions for authors

Q1. Could the authors further clarify how the theoretical assumptions relate to practical diffusion inverse solvers, particularly regarding the estimation of local prior uncertainty?

Q2. How accurately does the randomized-JVP estimator approximate the actual local prior covariance, rather than merely denoiser sensitivity?

Q3. Since reverse states depend on measurements, how is the assumed independence between local uncertainty and measurement noise justified?

Q4. Could the authors further distinguish TCO-based guidance from conventional covariance-weighted Kalman or Tikhonov corrections?

Q5. How is noise whitening implemented in the noiseless setting, given the assumption that the measurement covariance is positive definite?

Q6. Could the authors clarify whether the reported improvements generalize to alternative diffusion or flow-based backbones?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.