acceptodds
Under review as a conference paper at ICLR 2027

LOCUS: Feasible-Centered Null-Space Flow for Inverse Problems

Abstract

Inverse problems combine incomplete measurements with prior knowledge, from structural assumptions in compressed sensing to learned generative priors. For noiseless linear measurements , all measurement-consistent reconstructions lie in the affine fiber . Existing measurement-consistent constructions use this geometry to determine where generative transport may move, and raise a distinct design question: where within this feasible fiber should that transport be centered? We propose to predict an observation-dependent reconstruction, project it onto the feasible fiber, and use flow matching to generate the remaining null-space residual. This separates three roles: preserving what is measured, predicting what is inferable, and generating what remains after that prediction. The resulting dynamics preserve measurements throughout sampling. For independent projected-Gaussian sources and linear training paths, we show that excess center error relative to the conditional mean exactly equals excess displacement energy and ideal marginal-flow action. This gives prediction quality a direct interpretation in terms of residual transport. Across 17 large-cohort evaluations on FFHQ, CelebA-HQ, and ImageNet, spanning block and random-pixel inpainting and – super-resolution, the proposed design improves both PSNR and LPIPS over the corresponding pixel-centered null-space flow. A controlled 512-image study separates the center's conditioning and source roles, while a separate 2,000-image random-pixel benchmark demonstrates strong reconstruction and perceptual quality at 20 flow evaluations.

Primary area:
generative models
Keywords:
flow matching, inverse problems, image restoration, null-space generation, measurement consistency

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2k3wOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Locus is a conditional flow matching method, that first predicts a reconstruction, projects it onto an affine set that is measurement consistent, and then generates the remaining null-space residual of the forward operator. The resulting operation preserves noiseless linear measurements while using the learned prediction as the source center and also as a conditioning input. The authors theoretically relate center accuracy to displacement energy and ideal flow action, and empirically report improvements over a pixel-centered null-space flow on inpainting and super-resolution tasks.

Critical strengths
  • The decomposition of the inversion task in to: "preserve, predict, generate" is sensible, and conceptually strong.
  • Theory formalizes a previously heuristic design choice.
  • The ablation is through and design is excellent.
Critical weaknesses
  • Novelty: Paper is an extension of RN-SDE and Pixel-NFSM for flow priors.
  • Ablation shows very weak improvement supplying the learned center as network input, which weakens the story.
  • The scope of inverse problems is narrow. Noiseless linear measurements are a great start, but verifying results over ill-posed inverse problems with noise would have made the paper stronger and more sound.
  • No computational cost analysis.
Questions for authors
  • Please provide an equation-level comparison with RN-SDEs and clarify precisely which part of LOCUS and clarify which pars of LOCUS (beyond replacing the diffusion model with a flow matching model) is novel?
  • Could the authors report total training compute, parameter counts, wall-clock inference time, and FLOPs?
8 comments

Dear reviewer, I'm sorry to bother you.We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. If you are satisfied, we kindly request you to consider updating the score to reflect the newly added results and discussion. We remain committed to addressing any remaining points you may have during the discussion phase.

I thank the authors for clarifications on asked questions and formal distinction from previous approach. My main concern remains because source centering (which is the main point of the paper) produced only 0.012 dB gain in the original experiment. The new efficiency and noisy super resolution experiments sound very encouraging with interesting increase in LPIPS but PSNR suffers in more realistic conditions. The rebuttal establishes technical distinctness but does not yet demonstrate sufficiently strong novelty or broad empirical effectiveness.

Thank you for taking another look at the experiments. We agree that the 0.012 dB gain was small in the original one-step ablation, where both flows already received the learned center as input. Proposition 1 explains why a more accurate feasible center lowers the expected velocity-target energy. Our follow-up experiment tested whether changing the source also improves reconstruction at a fixed inference budget.

At the same inference budget, source centering improved PSNR by 0.884 dB in the controlled FFHQ block-inpainting experiment. We trained a fresh center and froze it for both flows. The flows shared the same architecture, initialization, conditioning inputs, and training budget. With 24 flow evaluations each, changing only the source mean raised PSNR from 29.31 to 30.19 dB. Both took about 384 ms per image, including the center call. The centered flow at 24 evaluations also outperformed the reference flow at 40 (30.19 vs. 29.39 dB) with 39.8% lower latency. The PSNR gap at equal budgets was negligible at 8 evaluations and ranged from 0.80 to 0.88 dB at the four tested budgets from 16 to 80. Across 17 large-cohort settings, the full LOCUS model improved both PSNR and LPIPS over Pixel-NSFM. That comparison includes both roles of the center.

The noisy super-resolution experiment measures what residual sampling adds once the measurement has been denoised. At a noise level of 15/255, the center alone achieved 28.63 dB. Adding the flow to the same center lowered PSNR to 28.22 dB while improving LPIPS from 0.213 to 0.160. We agree that this loss in pixel fidelity matters. Across all three noise levels, the center gave the higher PSNR, while the sampled reconstruction achieved better perceptual scores. Importantly, noisy measurements inherently contain less reliable information than their noise-free counterparts. Thus, higher PSNR should not be expected simply from introducing residual sampling into a noise-corrupted reconstruction; the relevant trade-off here is between pixel-wise fidelity and perceptual quality.

Thank you again for the thoughtful discussion and for pressing us to clarify where source centering makes a practical difference.

Reviewer_ZnyqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a novel end to end learning algorithm called LOCUS, where they learn a measurement-consistent center approximating E[X∣y,A]\mathbb{E}[X\mid y,A] and generate a plausible residual in the kernel, following the mantra “preserve what is observed and generate what remains.” The authors provide an interesting problem formulation to learn this null-space residual;since the center and the true signal satisfy the same measurements, their difference lies in the null space. Empirically, the results are competitive with methods using pretrained generative priors (which do not require task-specific training), particularly FLOWER, and improve perceptual similarity over baseline end to end methods.

Critical strengths

The key strength is the learning algorithm: combining a learned feasible center with null-space residual flow is interesting and novel. In addition, the paper is clear, and Figures 1 and 2 explain the method in detail. The results outperform several methods based on generative priors in perceptual similarity.

Critical weaknesses

The main weakness is that, although the proposed method consistently improves perceptual metrics across a wide range of tasks, the visual differences from some competing methods appear modest in Figure 4. These competing methods also do not require task-specific training, raising the question of whether the improvements justify the additional training cost.

Compared with the Feasible Center baseline, the residual flow improves perceptual quality but generally lowers PSNR. Although the paper acknowledges this tradeoff, a more detailed analysis of when and why residual generation is beneficial would strengthen the discussion.

Questions for authors
  • How does the trained model perform on unseen inpainting masks or different downsampling operators? Given the modest visual gains over pretrained generative priors, what justifies the additional training cost
  • Can the method control the tradeoff between the Feasible Center’s higher PSNR and LOCUS’s better perceptual similarity?
  • For the same measurement, do different noise samples produce meaningfully different, plausible reconstructions?
  • The conditional mean is a natural choice for the center. Did you explore other estimators, and was training stability a factor in your choice?
4 comments

Dear reviewer, I'm sorry to bother you.We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. If you are satisfied, we kindly request you to consider updating the score to reflect the newly added results and discussion. We remain committed to addressing any remaining points you may have during the discussion phase.

I would like to thank the authors for there detailed rebuttal, I am going to raise my score from 2 to a 3.

Thank you again for taking the time to discuss these questions with us in such detail, and for raising your score. Your constructive comments have substantially improved the quality of our manuscript. We really appreciate it.

Thank you very much for your thoughtful feedback. We sincerely appreciate your support for the acceptance of our paper and are grateful that our response effectively addressed your concerns.

But we noticed that while you mentioned increasing your rating, the updated score might not have been reflected in the system yet. And we just wanted to check in case it was an oversight.

We truly appreciate your valuable comments and support.We would like to express our sincere thanks to you again.

Reviewer_2PpmOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LOCUS, a two-stage, task-trained method for noiseless linear inverse problems. In the first stage, an image predictor is trained with an MSE loss and its output is projected onto the feasible set {z:Az=y}\{z : Az = y\}, giving a measurement-consistent "center" c(y)=A†y+PGϕ(hA(y))c(y) = A^\dagger y + P G_\phi(h_A(y)), i.e., an initialization of the null-space component. In the second stage, a flow-matching model learns the remaining null-space residual: the source is c(y)c(y) plus projected Gaussian noise, the velocity is projected onto ker⁡(A)\ker(A), and the center is also given to the network as a conditioning input. This is based on a data-dependent coupling between the center and the clean image, and every state along the trajectory is consistent with the measurement. Proposition 1 shows that the excess center error relative to the conditional mean equals the excess velocity-target energy and ideal flow action. Experiments on FFHQ, CelebA-HQ and ImageNet (block and random-pixel inpainting, 4×–16× SR) show consistent improvements over a pixel-centered null-space flow (Pixel-NSFM), and better perceptual metrics than zero-shot solvers and PMRF.

Critical strengths
  1. I find the idea of learning the residual/fine details via a data-dependent coupling interesting, and the projected construction guarantees measurement consistency by design.
  2. The experimental evaluation is broad: three datasets, six tasks, a wide range of baselines, and large test cohorts. The controlled B/F/C study (Sec. 4.3) is a useful attempt to separate the two roles of the center.
Critical weaknesses
  1. Missing related work. Several closely related works are not discussed:

    • Sampling the residual: Whang et al. [1] already propose a predict-and-refine approach, where a deterministic predictor is followed by a diffusion model that samples the residual. This is very close to LOCUS and should be discussed, and ideally compared against.
    • Data-dependent couplings and conditional generation from the measurement, which belong in the paragraph "Predictive sources and null-space generation": stochastic interpolants with data-dependent couplings [2], InDI [3] and Palette [4].
    • Operator-aware models: multitask learning with stochastic interpolants [5], which trains a single interpolant-based model across multiple operators; energy models learned via anisotropic diffusion [6]; and degradation-aware models such as InvFusion [7] and RAM [8]. These are relevant because LOCUS is trained separately for each operator.
  2. Why is learning the residual better? As the related-work section itself states, LOCUS combines PMRF (flow from a predicted posterior mean plus noise) with a null-space flow (NullFlow / Pixel-NSFM). Proposition 1 is essentially a bias–variance decomposition: a better center reduces the velocity-target energy, but, as the authors acknowledge in Appendix A, this does not imply better samples from a finite network. The paper's own controlled study suggests that source centering, the main thesis of the paper, contributes little (Sec 4.3). Overall, I don't think the paper does a good job of explaining why centering the transport at the prediction is better than methods like PMRF, or than simply conditioning the flow on a good predictor. The comparison with PMRF is also limited to one task (random-pixel inpainting) with different architectures. Without a clearer distinction from these methods, and a proper discussion of [1], the novelty feels incremental.

Questions for authors
  1. The method requires two networks, the image predictor and the flow model, both trained per operator. The center's training is excluded from the comparison with Pixel-NSFM ("Center training is additional"), and Appendix B.4 states that the LOCUS/Pixel-NSFM pairs "are not common-initialization or equal-total-training-compute experiments". How does LOCUS compare with a Pixel-NSFM baseline with matched total parameters or training compute? This would show whether the gains (0.11–0.40 dB) come from the extra capacity.
  2. Table 2 compares a task-trained model with zero-shot solvers (DDNM, FlowDPS, PnP-Flow, FLOWER) that use a single prior for all tasks. Could the authors state this difference clearly in the main text?
  3. Could the authors report the training details of the center predictor (architecture, number of parameters, optimizer, learning rate and number of updates per task)? Table 4 only lists the flow-stage settings, and the checkpoint history (inherited 4× checkpoints, resumed runs, 5k adaptation stages) makes the results hard to reproduce.
  4. The construction relies on the hard constraint Az=yAz = y, so it does not handle measurement noise (App. F.1), which is the standard setting for most of the zero-shot baselines. How would the method extend to noisy measurements? A small experiment would strengthen the paper.
  5. Could the authors add a pseudocode summarizing the two training stages and the sampler? The description is currently split between Sec. 3 and Appendices B.2–B.3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.