acceptodds
Under review as a conference paper at ICLR 2027

TubeFM: Likelihood-Isometric Flow Matching for Noisy Inverse Problems

Abstract

Posterior samplers for noisy inverse problems must represent uncertainty in measured as well as unmeasured directions. TubeFM learns conditional flows in an affine chart comprising a fiber coordinate and a noise-standardized measurement residual. We use this chart to connect posterior approximation, velocity regression, and numerical integration. An exact Wasserstein decomposition quantifies the error imposed by hard measurement constraints; a flow-stability bound transfers velocity error to image space; and explicit prior assumptions yield noise-uniform posterior curvature bounds. Gaussian analysis separates this target geometry from the dynamics of the chosen probability path. With the source, path, and error criterion transformed together, an affine change of coordinates preserves the mapped Runge–Kutta trajectory and evaluation count in exact arithmetic; changing the source can alter both. Experiments on 135 Gaussian problems, real handwritten images, procedural Fourier imaging, and an exploratory fastMRI pilot test these distinctions. Three-seed learned-flow comparisons show that a source-matched image model can outperform TubeFM in reconstruction, while hard projection removes measured-coordinate uncertainty. The MRI pilot reproduces this uncertainty distinction with nearly unchanged PSNR on two validation volumes. We also quantify the finite-ensemble effect in raw percentile coverage. Together, the analysis and controls identify what likelihood scaling contributes and separate it from source selection, posterior calibration, and solver efficiency.

Primary area:
generative models
Keywords:
Conditional flow matching, Bayesian inverse problems, posterior sampling, uncertainty quantification, likelihood geometry

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_smL3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TubeFM, a conditional flow-matching method for noisy linear inverse problems. The main idea is to work in coordinates that separate the null-space, or fiber, component of the signal from a noise-normalized residual in the measured subspace. The paper develops several theoretical results around this construction. These include an exact decomposition of the Wasserstein error introduced by restricting samples to a hard measurement slice, a stability bound relating velocity approximation error to endpoint distribution error, and Gaussian calculations concerning posterior curvature and flow dynamics. The paper also emphasizes that changing coordinates, changing the source distribution, and changing the numerical error metric are different interventions. The experiments cover Gaussian problems, low-resolution handwritten digits, procedural Fourier imaging, and a small fastMRI pilot. The results are mixed. TubeFM preserves uncertainty that is removed by hard projection, but the source-matched image-space CFM obtains better reconstruction PSNR than TubeFM in the learned digit experiment.

Critical strengths
  1. The Gaussian control experiments are carefully designed. In particular, the authors show that changing coordinates alone does not reduce NFE when the source, path, and error metric are matched. This helps clarify where the computational differences actually come from.

  2. The theoretical analysis provides a useful connection between likelihood geometry, flow approximation error, and posterior uncertainty. The Wasserstein decomposition also makes the uncertainty lost through hard projection explicit.

  3. The paper is transparent about its negative results and limitations. The authors report that the source-matched image CFM outperforms TubeFM in reconstruction and clearly distinguish reconstruction quality from uncertainty calibration.

Critical weaknesses
  1. (Novelty) The core construction combines likelihood whitening, SVD-based null-space decomposition, and standard conditional flow matching. These ingredients are well established, and several theoretical results follow from standard arguments. It is difficult to identify a sufficiently strong methodological contribution beyond applying this coordinate system to CFM.

  2. (Significance) In Table 2, the source-matched image CFM achieves higher PSNR than TubeFM at all three noise levels. This suggests that the source distribution may be more important than the proposed coordinates. The practical benefit of learning in tube coordinates remains unclear.

  3. (Ablation) The learned comparison does not fully isolate the effect of the coordinate transformation. TubeFM and the matched-source image baseline use different regression metrics, so their performance differences may also come from loss weighting rather than the coordinates themselves.

  4. (Experiments) The learned experiments are quite limited. The main comparison uses 8x8 digit images with a fixed observation mask, while the MRI pilot evaluates only two validation volumes and one training seed. This makes it difficult to assess whether the findings generalize to realistic inverse problems.

  5. (Applicability) The method relies on a spectral factorization or null-space representation of the observation operator. It is unclear how efficiently the proposed coordinates can be constructed for large implicit operators where an explicit SVD is impractical.

Questions for authors
  1. Given that the source-matched image CFM consistently achieves higher PSNR in Table 2, what is the main practical advantage of using TubeFM instead of simply using the same source distribution in image coordinates?

  2. Could the authors separate the effects of coordinate transformation and regression loss? For example, comparing image CFM with the pulled-back tube loss and TubeFM with an image-space loss would help identify which component is responsible for the observed differences.

  3. Could the authors provide additional learned experiments on higher-resolution images or more MRI volumes, preferably including a source-matched image baseline and multiple training seeds?

  4. The learned experiments mainly report reconstruction and coverage metrics, which do not necessarily establish posterior accuracy. Could the authors evaluate learned TubeFM on a problem with a known posterior to directly measure the approximation error?

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

The paper explores posterior approximation for linear inverse problems with Gaussian noise. The authors argue that enforcing exact data consistency by replacing measured components with the observed values is incorrect when we have noisy measurements, because it removes the uncertainty in the measured components. They propose TubeFM, a conditional flow-matching model trained in transformed coordinates that separate the unmeasured part of the image from the measurement residual. Their experiments show that TubeFM preserves uncertainty in the measured components. However, an image-space baseline achieves better reconstruction quality in Table 2.

Critical strengths

None

Critical weaknesses
  1. The paper is very hard to follow. I spent considerable time understanding the main idea, but I still could not follow most of the paper. The paper uses several non-standard or ambiguous terms, which makes it harder to follow; for example, “partially observed inverse problem”, “positive noise”, “noisy likelihood”, and even “fiber”. Fiber is a term from topology and geometry and is not common in the imaging community; the standard phrasing would be “the set of measurement-consistent solutions”. I recommend using standard inverse-problems vocabulary and defining any new terms explicitly. In addition, please simplify the language of the paper to make it more readable.

  2. The paper’s motivation is on the drawbacks of hard data consistency under noisy measurements. However, many established diffusion-based solvers like DPS [1], DDRM [2] and DiffPIR [3] already incorporate the noise model without hard projection. The paper discusses some of these in related work but does not compare against them experimentally.

  3. The experiments are limited to very low-dimensional settings 8×8, 32×32. The only high-dimensional experiment is an MRI pilot: one. It is therefore unclear whether the paper’s findings, including the comparison with the matched-source baseline, hold at realistic image resolutions or on standard benchmarks used by related diffusion-based methods.

[1] Chung, Hyungjin, et al. "Diffusion posterior sampling for general noisy inverse problems." ICLR 2023.

[2] Kawar, Bahjat, et al. "Denoising diffusion restoration models." NeurIPS 2022.

[3] Zhu, Yuanzhi, et al. "Denoising diffusion models for plug-and-play image restoration." 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 2023.

Questions for authors

None

Reviewer_khPQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a method to sample from the posterior for linear inverse problems. It introduces an invertible coordinate transform into a new “tube” coordinate system, which separates the measured and unmeasured directions. Experiments are done on gaussian toy tasks, handwritten digits, synthetic MRI phantoms, and MRI data. Their experiments show that the coordinate change does not change sampling cost and that the image space flow with the same source distribution beats TubeFM on PSNR. Given those results the contribution appears to be more in the analysis than a superior methodology. Unfortunately, the data they tested on is relatively small scale.

Critical strengths
  1. a well motivated reparametrization into “tube” coordinates which is methodologically close to DDRM and conditional flow matching for posterior estimation.
  2. a clean analysis, which decouples effects of the chosen coordinate transform and the source distribution.
Critical weaknesses
  1. Applicability: The proposed method only supports linear inverse problems and Gaussian noise. Furthermore, it requires a well known SVD of Σ−1/2 A, Σ and A must be known at training and test time. Wide application of this method is therefore not directly possible.
  2. Novelty is limited: DDRM already works in the SVD basis, DDRM is not compared against. Theorem 1 is a well known fact it is only relevant for methods which enforce hard data consistency (HDC). It would be good to clarify which ones those are. Usually trained CFM methods do not enforce HDC.
  3. Evidence: no convincing evidence that the method beats competitive approaches, experiments are mostly on toy tasks and small networks.

Minor Points: The paper is difficult to read and to follow. E.g.:

  1. The terminology of tube, fibre, least squares slice etc. are used before they are explained.
  2. Theorems and coordinate transform are stated without stating their motivation of what question they answer.
  3. It is a bit difficult to follow what the actual contribution is when one starts reading the paper.
  4. a better differentiation between mentioned plug and play approaches like DPS, DDRM and so on versus trained CFM.
  5. While cautious claims are fair and limitations should always be mentioned there are a lot of caveats and exceptions in the text which make it difficult to follow the point. Who actually makes the "curvature determines solver cost" argument?
Questions for authors

Suggestions: The paper starts by proposing a new method "TubeFM". Throughout the methods and experiments the paper analyses that the proposed method does not improve metrics e.g. to source matched image flow. In the end the fair analysis within the paper gives more insights than the method includes novelty or proves applicability. Potentially the paper is better posed as an analysis paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.