acceptodds
Under review as a conference paper at ICLR 2027

Mean Flows as Controllable Priors for Solving Inverse Problems

10 views

Abstract

Training-free inverse solvers built on flow priors repeatedly need to predict the clean sample reached from an intermediate state along the generative trajectory. Standard flow-matching models expose only an instantaneous velocity, so such predictions need numerical integration or coarse extrapolation. MeanFlow models learn a finite-interval transport map, which lets us use them as controllable priors that move a state across any interval, to the next stage or clean endpoint—all in one evaluation. This interface carries the structure of any flow-based solver onto a MeanFlow prior and makes a new control policy natural: Receding-Horizon State Optimization (RHSO), a model-predictive controller whose decision variable is the generative state itself. A first-order analysis for general data distributions shows how distributing optimization relaxes but retains the prior's implicit regularization. Capacity-matched comparisons in pixel and latent space measure what MeanFlow priors contribute: they surpass flow-matching priors at every stage count tested, most widely with few stages and in latent space. Across five ImageNet restoration tasks, four solvers and four priors, RHSO is best in 35 of 60 task–metric cells and wins 200 of 240 pairwise comparisons. A depth-controlled study attributes most of its gain to distributing optimization over stages, not deeper numerical planning.

Primary area:
generative models
Keywords:
inverse problems, flow matching, MeanFlow, generative priors, training-free solvers, model predictive control, image restoration

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_WhU5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes the use of MeanFlow models as generative priors for solving inverse problems. MeanFlow models are a type of generative model capable of transporting an intermediate sample from time s to time t, for any s and t, and have the key benefit of avoiding the cost of long rollouts. The manuscript proposes and analyses a control policy, Receding-Horizon State Optimization, that introduces measurement guidance to a MeanFlow-based flow solver. These contributions are evaluated against existing flow-based inverse problem algorithms with flow matching and MeanFlow priors across different linear image-based inverse problem tasks.

Critical strengths

Cheap and accurate estimation of the endpoint is important for guiding generative flows, and as a result, MeanFlows are a very natural and interesting direction to explore. In particular, the experiments comparing flow matching models to MeanFlows and the ablations are scientifically interesting and would have applicability to related work on guiding pre-trained generative models.

Critical weaknesses
  1. The evaluation is limited to a single dataset, and so it is unclear whether the conclusions will generalise further. In particular, the experiments jump straight to ImageNet and million-parameter models. It would be more instructive to see a controlled comparison of the proposed approach in a simpler setting with a known (e.g., pre-specified and low-dimensional) prior for validation. This would also address most of the limitations stated in the limitations setting, such as comparison with matched data, architecture and training recipe.

  2. The relative novelty appears somewhat incremental over related methods, apart from the replacement of flow models with MeanFlows. Indeed, the RHSO policy seems to be a re-parameterisation of MPC-$\Delta t. The improvement in quantitative metrics is also not separated from increased computational cost.

Questions for authors
  1. [Clarification] Equation (13) appears to be mathematically equivalent to MPC-Δt\Delta t under a simple reparameterisation:

In MPC-Δt\Delta t, let uu denote the optimised control and define the uncontrolled flow step as

s=xtk−1+Δtvθ(xtk−1,tk−1).s=x_{t_{k-1}}+\Delta t v_\theta(x_{t_{k-1}},t_{k-1}).

Making the substitution

x=s+Δtu,u=x−sΔtx=s+\Delta t u,\qquad u=\frac{x-s}{\Delta t}

transforms the MPC-Δt\Delta t objective into

\Phi(P_k(x);y) +\frac{1}{\lambda\Delta t^2}\|x-s\|_2^2 \right].$$ This is identical to Eq. (13) when $$\boxed{\mu=\frac{2d}{\lambda\Delta t^2}}$$. Could the authors clarify how their formulation differs from MPC-$\Delta t$, beyond this change of variables? 2. [Future work/clarification] How does RHSO perform at matched computational cost to the baseline algorithms? At present, its improvement in quantitative metrics may be attributed to the increased computational budget it is allocated (see line 370) compared to the baselines. 3. [Clarification] RHSO could be more clearly motivated; Section 3.2 says that “those two properties make a specific controller natural”, but I don’t see why the choice of equation (13) is natural/unique from the preceding section? 4. [Clarification] In Step 5 of Algorithm 1, why is the proximal anchor weight a constant and not dependent on the timepoint (c.f. FLOWER)? 5. [Clarification] Why compare only to 1-step D-Flow? Why not multi-step? 6. [Future work] Can you add FLOWER to the comparison? While I understand and appreciate the desire to avoid a leaderboard of methods, I think that it would be interesting to see how FLOWER performs compared to your method as it shares many characteristics such as the alternating proximal update structure. 7. [Future work/rewriting] In places, the paper is hard to follow. For example, lines 373-377 feature sentences with many clauses that require a lot of unpacking from the reader, while line 377-378 is an example of a statement that could be restructured to make the reader’s journey through the prose cleaner: “Figure 2 shows the behaviours behind the numbers.” (These are just illustrative examples; other sections would also benefit from simpler sentence structures and less abbreviated descriptions.)
4 comments
AuthorsOpenReview ↗

Response to WhU5 (1/2): relation to MPC-Δt, matched compute, motivation of Eq. (13)

We thank the reviewer for their careful and detailed reading of our paper. Working through your questions led to new experiments and revisions that we believe make the paper stronger. We answer each point below and list the corresponding changes. Section, table and appendix numbers refer to the revised paper.

1. Relation of Eq. (13) to MPC-Δt. Thank you for working this out; the change of variables is correct. The difference lies in the fact that MPC-Flow controls the generative ODE additively, x˙=vθ(x,t)+u(t)\dot x = v_\theta(x,t) + u(t): the control is added to the network's output, after the network has been evaluated, not to its input. For MPC-Δt, one stage therefore reads

xtk+1=xˉ+Δu,xˉ=xtk+Δ vθ(xtk,tk),x_{t_{k+1}} = \bar x + \Delta u, \qquad \bar x = x_{t_k} + \Delta\, v_\theta(x_{t_k}, t_k),

where xˉ\bar x does not depend on uu (for a MeanFlow prior, xˉ=Tθ(xtk;tk→tk+1)\bar x = T_\theta(x_{t_k}; t_k \to t_{k+1})). RHSO instead moves the state before the network transports it, xtk+1=Ek(xtk+δk)x_{t_{k+1}} = E_k(x_{t_k} + \delta_k), so its control is an input to the network rather than an increment added to its output. Substituting x′=xˉ+Δux' = \bar x + \Delta u turns the MPC-Δt stage objective into Eq. (13) at time tk+1t_{k+1}, with μ=2d/(λmΔ2)\mu = 2d/(\lambda m \Delta^2), but the correction still acts at a different place relative to the network. The case N=1N = 1 makes this explicit. MPC-Δt samples xˉ=x0+vθ(x0,0)\bar x = x_0 + v_\theta(x_0, 0) and then solves min⁡u∥u∥2+λ Φ(xˉ+u)\min_u \|u\|^2 + \lambda\,\Phi(\bar x + u), a regularized correction in image space that never backpropagates through the prior. RHSO solves min⁡x0Φ(Tθ(x0;0→1))+μ2d∥x0−s0∥2\min_{x_0} \Phi(T_\theta(x_0; 0 \to 1)) + \frac{\mu}{2d}\|x_0 - s_0\|^2 from the sampled noise s0s_0, so every update passes through the network and the output stays in the range of the prior map. This difference is visible in practice, both at equal cost (point 2) and in the new 2-D study (below).

Changes: Section 3.2 (paragraph "Relation to D-Flow and MPC-Flow") states this contrast; Appendix A.1 gives the change of variables.

2. Matched computational cost. We added an equal-cost comparison on pMF-L/16 (Section 4.4, Table 3; details in Appendix K). When every other method's budget is raised to RHSO's wall-clock time, RHSO still has the best five-task mean in every metric and is better in 65 of 75 comparisons. When RHSO is instead cut to MPC-Δt's backpropagation budget, it is still better than MPC-Δt in 12 of 15 comparisons (Table 33). Only at the much lower budgets of PnP-Flow and FLOWER, which do not backpropagate through the prior, are these two methods stronger; we state this in the limitations. We also rewrote the description of RHSO's cost that you pointed to (line 370), now at the start of Section 4.4, with time and memory for every method and prior in Appendix I.

3. Motivation of Eq. (13). We do not claim that Eq. (13) is uniquely determined by the MeanFlow interface; "natural" referred to cost. A finite-interval map gives a terminal prediction from any stage for one evaluation and one backpropagation, so the state itself can be optimized at every stage. With a flow-matching prior, the same objective needs either an integrated suffix whose cost grows with the remaining horizon, or a one-step extrapolation. Eq. (13) is a direct and simple receding-horizon controller built on this, with no additional control variables, and it performs well empirically (Sections 4.3–4.6). We revised the corresponding sentence in Section 3.2.

AuthorsOpenReview ↗

Response to WhU5 (2/2): anchor, multi-step D-Flow, FLOWER, writing, low-dimensional study, novelty

4. Time-dependent anchor. We compared the constant anchor with two schedules μk=μ0 m(tk)\mu_k = \mu_0\, m(t_k) on pMF-L/16, each with its own tuned μ0\mu_0. The first follows FLOWER's schedule: with FLOWER's νt=(1−t)/t2+(1−t)2\nu_t = (1-t)/\sqrt{t^2+(1-t)^2}, it uses m(t)=νt−2=(t2+(1−t)2)/(1−t)2m(t) = \nu_t^{-2} = (t^2+(1-t)^2)/(1-t)^2, which increases with tt. The second uses its reciprocal, m(t)=νt2m(t) = \nu_t^{2}, which decreases. Neither improves on the constant anchor: the increasing schedule is worse in 11 of 15 task–metric comparisons and better in 1, and the decreasing one gains at most 0.03 dB and 0.005 LPIPS where it is better while losing 1.6 dB on denoising. We therefore keep the constant anchor (Appendix H.2, mentioned in Section 4.6).

5. Multi-step D-Flow. The multi-step version is in Table 4 (Section 4.6). On JiT-B/16, with the same 160 source-state updates, integrating D-Flow's terminal prediction over four Heun intervals raises it from 18.18 to 20.84 dB at seven times the cost, while RHSO, which distributes the same updates over the trajectory, reaches 28.47 dB at equal cost. The benchmark uses the one-step form because it matches the other solvers' cost per update.

6. FLOWER. We added FLOWER on both pixel-space priors, with its step count tuned by the same protocol as every other solver (Section 4.3; Tables 1, 3, 8, 9 and 12). FLOWER and PnP-Flow are close (FLOWER is better in 7 of 15 comparisons on pMF-L/16 and 8 of 15 on JiT-B/16). RHSO is better than FLOWER in 12 of 15 comparisons on pMF-L/16 and 9 of 15 on JiT-B/16; with FLOWER included, it is best in 32 of 60 task–metric cells and wins 221 of 270 pairwise comparisons. We updated the abstract and Section 4.3 accordingly.

7. Writing. Thank you for the concrete examples. We rewrote lines 373–378 into short sentences with one idea each, moved their discussion of cost to Section 4.4, and replaced the Figure 2 sentence with a direct one. We also restructured Section 4 into a short introduction, the 2-D study and the ImageNet experiments, each subsection answering one question.

Controlled low-dimensional study. Following your suggestion, Section 4 now opens with a controlled 2-D study before the ImageNet experiments (Section 4.1, Appendix L). Flow-matching, MeanFlow and improved-MeanFlow priors are trained on a hexagon with the same data, network and recipe, and every solver runs on all three. Two findings carry over to images. One-step D-Flow fails on the flow-matching prior but works on both MeanFlow priors. And where the correction acts matters: steered toward an off-data target, MPC-Δt's output on the MeanFlow priors stops about a third of the way from the data to the target, while RHSO's stays near the data, closer to the nearest data point than MPC-Δt's on every prior. PnP-Flow and FLOWER are as accurate as RHSO there.

Novelty. We see the main contribution in the prior. A pretrained MeanFlow map gives the clean endpoint from any state in one evaluation, with a one-backpropagation gradient whatever the remaining horizon, so existing solvers carry over to it unchanged (Section 3.1), and at matched capacity and cost MeanFlow priors are better in 153 of 165 task–metric cells in pixel space and all 165 in latent space (Section 4.5). RHSO is a simple controller built on this interface; it differs from MPC-Δt in where the correction acts (point 1), and this matters in practice (point 2 and the 2-D study).

We hope these changes address the reviewer's concerns, and we would be grateful if the reviewer would take them into account in the final assessment.

Thank you for your detailed yet concise reply.

The revisions to the manuscript make clear the difference with MPC-Flow; thank you for clarifying my misunderstanding here. The rest of the answers to the questions I posed are convincingly answered.

As this rebuttal addresses W2 convincingly and makes substantial progress towards addressing W1, I will consider revising my scores upwards.

Thank you for engaging with our response and for the constructive review. We are glad the revisions answer your questions.

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

This paper proposes an algorithm that can leverage a pretrained Meanflow model or a flow matching model to solve image inverse problems. The proposed algorithm, Receding Horizon State Optimization (RHSO), is an MPC-style N-step iteration algorithm, that at each timestep t estimates the clean image and then updates the current noisy image with M steps of gradient descent, and then uses this to estimate the next state xt+1x_{t+1}. The pretrained models can be leveraged in estimating the clean image. The algorithm has been evaluated on standard image inverse tasks such as gaussian deblurring, denoising, 2x super resolution, random inpainting, and box inpainting. Standard metrics such as PSNR, SSIM, LPIPS have been reported in addition to Measurement RMSE. The method has been evaluated in both pixel space and latent space and on both meanflow and flow matching.

Critical strengths
  1. This paper provides a way to leverage Meanflow in prior methods such as D-Flow, PnP-Flow, MPC-RHC and MPC-Δt\Delta t, while also proposing another algorithm that outperforms them. Meanflow also allows lightweight backdrop from time t without any ODE unrolling.
  2. Empirical benchmarking is quite broad across both latent and pixel space, 5 different inverse problems and for both Meanflow and FlowMatching model. In addition, the appendix includes ablations of sweeps of various hyperparameters of RHSO.
Critical weaknesses
  1. Gap between empirical and theoretical claims on why RHSO works: In section 3.3, the paper claims that the Jacobian Jt=DxΨt→1(xt)J_t = D_x \Psi_{t \rightarrow 1}(x_t) regularizes the image by penalizing directions the remaining transport barely amplifies. These directions would be off-manifold for images, but the proof of proportion proof excludes low dimensional manifolds.
  2. RHSO is computationally quite expensive. As shown in Table 6, For SiT-XL architecture, RHSO uses 5776 model evaluations and 5760 back-rope for all image restoration tasks. In contrast, the other baselines use substantially less compute! The paper skips wall clock times for SiT, and also, peak memory requirements for all baselines and RHSO.
  3. Appendix B that includes theory and proofs is not very well presented and a bit difficult to parse. It would be useful to spend some time rewriting this section so that the readers can more readily understand the core theoretical results.
Questions for authors
  1. Could the authors provide peak memory requirements of the method and the baselines?
2 comments
AuthorsOpenReview ↗

Response to Reviewer rSAb: near-manifold theory, timing and memory, Appendix B

We thank the reviewer for the positive assessment and for the specific, constructive points, which led to a new theoretical result, complete timing and memory measurements, and a clearer Appendix B. We answer each point below. Section, table and appendix numbers refer to the revised paper.

1. Theory and low-dimensional manifolds. Thank you for pointing this out. The late-stage result (Proposition 4) requires bounded curvature (condition (27)), which excludes data lying exactly on a low-dimensional set (Remark 5). We added Corollary 7, which shows that this condition holds for data arbitrarily close to such a set: if data on a manifold is smoothed by Gaussian noise of any variance ε2>0\varepsilon^2 > 0, condition (27) holds with explicit constants, and all conclusions of Proposition 4 apply. The pricing of corrections (Proposition 2) needs no such condition, and for data exactly on a manifold it makes off-manifold corrections unreachable at first order. Together, the intuition of Section 3.3 holds for data arbitrarily close to a low-dimensional manifold, a common model of natural images, and the 2-D study of Section 4.1 (added in our revision), whose data lies on a curve, agrees: RHSO's outputs stay on the data. Section 3.3 now states this, and Appendix B opens with an overview of which result needs which assumption.

2. Computational cost and memory (Weakness 2 and the question). Thank you. In our previous revision, we added wall-clock time and peak memory for every method on all four priors, SiT-XL/2 included (Appendix I, Tables 31–32).

You are right that RHSO is expensive on SiT-XL/2. On a flow-matching prior, its terminal prediction needs an integrated suffix, so each update differentiates through several network evaluations; this makes it the most expensive solver there. On a MeanFlow prior, the same prediction is a single evaluation. On iMF-B/2, our latent-space MeanFlow prior, RHSO's runtime is on par with D-Flow (1-step). On pMF-L/16 it is faster than D-Flow (1-step) on every task and faster than MPC-Δt on four of five, with the same peak memory as both (Table 32). This gap between SiT-XL/2 and the MeanFlow priors is the practical benefit that the paper argues for.

Our previous revision also added an equal-cost comparison of all methods (Section 4.4). With every other method given RHSO's wall-clock time on pMF-L/16, RHSO has the best five-task mean in every metric. At much lower budgets, PnP-Flow and FLOWER (a baseline added in our previous revision), which do not backpropagate through the prior, are stronger, as stated in the limitations.

3. Presentation of Appendix B. Thank you for the suggestion. We restructured Appendix B: it now opens with an overview of the results, what each assumes and where the main text uses it; each result is preceded by a short plain-language statement of what it shows; and the longer proofs are split into labelled steps.

We hope these changes address the reviewer's points, and we thank the reviewer again for the helpful review.

Thank you for the response and updating the manuscript. My concerns have been addressed while some minor ones remain. I am retaining my score for now.

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

This paper focuses on the use of pretrained flow-matching models as priors for the solution of inverse problems. They propose to replace standard flow-matching models with MeanFlow models, which transport a state between arbitrary timesteps in a single step. The authors show that the MeanFlow priors adapt to various existing inverse-problem solvers by replacing their endpoint prediction with the MeanFlow map. They also propose a new solver called Receding-Horizon State Optimization (RHSO), which is a multi-stage version of an existing solver (D-Flow). Specifically, instead of optimizing only the initial noise, their method repeatedly optimizes the intermediate states to improve measurement consistency, and then proceeds to the next time step using the flow map. The method applies to both MeanFlow and standard flow-matching models, although MeanFlow makes endpoint predictions computationally cheaper. They show improvements over their chosen baselines in multiple image inverse problems.

Critical strengths

The paper is well-organized, and the direction of using MeanFlows as generative priors for inverse problems is worthy.

The algorithm is simple and easy to use.

The experimental setup is sound and includes different ablations.

Critical weaknesses

The computational cost of RHSO is high. Despite reducing the cost of endpoint prediction, the method still requires many backpropagations through the flow network.

The novelty is limited, as the method combines existing methods without a clear probabilistic view on the posterior sampling.

Some sentences are too long and repetitive. I would suggest to convey one idea per sentence. I believe the presentation of the theoretical section could improve.

Questions for authors

I would be interested in benchmarking other flow map learning paradigms versus MeanFlow as one-step generative solvers i.e. distillation-based, consistency models.

What is the intuition or suggestion for setting the mu parameter (weight of anchoring)?

Can the method be used for blind inverse problems where the forward operator is not known?

2 comments
AuthorsOpenReview ↗

Response to Reviewer yJUX (1/2): computational cost, novelty and probabilistic view

We thank the reviewer for the positive assessment and the thoughtful questions. Your point on computational cost motivated the equal-cost comparison and the time and memory measurements added in the revision. We answer each point below. Section, table and appendix numbers refer to the revised paper.

1. Computational cost. The revision compares all methods at equal cost on pMF-L/16 (Section 4.4, Table 3; details in Appendix K). When every other method is given RHSO's wall-clock time, RHSO has the best five-task mean in every metric and is better in 65 of 75 comparisons; when RHSO is instead cut to MPC-Δt's backpropagation budget, it is still better than MPC-Δt in 12 of 15 comparisons. At much lower budgets, PnP-Flow and FLOWER (a baseline added in the revision) are stronger, as stated in the limitations.

The revision also reports wall-clock time and peak memory for every method on all four priors (Appendix I, Tables 31–32). They show two tiers. Solvers that backpropagate through the prior (D-Flow, MPC-Δt and RHSO) share the slower tier, and PnP-Flow and FLOWER, which do not, are markedly faster. Within the first tier, RHSO is not the most expensive on MeanFlow priors: on pMF-L/16 it is faster than D-Flow (1-step) on every task and faster than MPC-Δt on four of five, with the same peak memory, and on iMF-B/2 its runtime is on par with D-Flow (1-step).

2. Novelty and the probabilistic view. We agree that RHSO builds on existing ideas: it distributes D-Flow's source optimization over the trajectory in the receding-horizon manner of MPC. The main novelty lies in replacing flow-matching priors with MeanFlow priors. Their finite-interval map gives the terminal prediction and its gradient from any intermediate state with one evaluation and one backpropagation, so existing solvers carry over unchanged and every stage of RHSO can be optimized through the map to the clean image. The capacity-matched study of Section 4.5 measures this benefit: the MeanFlow prior is better in 153 of 165 task–metric cells in pixel space and in all 165 in latent space.

RHSO is an optimization-based solver, like D-Flow and MPC-Flow, and its objective has a direct probabilistic reading. For a deterministic generative ODE, the image is a function of any intermediate state, x1=Ψt→1(xt)x_1 = \Psi_{t\to1}(x_t), so the likelihood of a state is available exactly: p(y∣xt)=p(y∣Ψt→1(xt))p(y \mid x_t) = p(y \mid \Psi_{t\to1}(x_t)). Its gradient is the endpoint likelihood gradient pulled back through the transport Jacobian JtJ_t, which is the likelihood term of the score of the state posterior p(xt∣y)∝pt(xt) p(y∣xt)p(x_t \mid y) \propto p_t(x_t)\,p(y \mid x_t); a MeanFlow prior evaluates it with one evaluation and one backpropagation. The data term of Eq. (13) is this log-likelihood up to the weighting set by μ, and the anchor is a local quadratic regularizer around the state reached by the preceding stage, standing in for ptp_t; with μ = 0, the prior acts only through the transport geometry analysed in Section 3.3.

Empirically, the 2-D study (Section 4.1, Appendix L, added in the revision) includes an inverse problem with a known posterior. Across random initial noise, RHSO's outputs are almost as close to exact posterior samples (sliced W2 0.51–0.52) as a second set of exact samples is (0.50).

AuthorsOpenReview ↗

Response to Reviewer yJUX (2/2): writing, flow maps, anchor weight, blind problems

3. Writing and the theory section. Thank you. The revision restructures Section 4 and rewrites its densest passages in shorter sentences, one idea each. It also restructures the theory in Appendix B: an overview of the results, what each assumes and where the main text uses it; a short plain-language statement before each result; and the longer proofs split into labelled steps.

4. Other flow-map paradigms. Thank you for this suggestion. RHSO needs a terminal prediction from each stage and a transport of the state to the next stage. Distilled flow maps such as Consistency Trajectory Models (Kim et al., 2024) and Align Your Flow (Sabour et al., 2025) learn jumps between arbitrary times along the trajectory, like MeanFlow, so RHSO and the other solvers apply to them unchanged through the interface of Section 3.1. Standard consistency models (Song et al., 2023) also provide the terminal prediction, and the transport can be realized by their multistep procedure, which predicts the clean endpoint and re-noises it to the next time. RHSO accommodates this, but the transport is then a stochastic re-noising step rather than a jump along the trajectory, so it is a distinct variant of the method rather than a change of prior alone. A fair comparison of these families as priors needs models matched in data, architecture and training, as our capacity-matched study provides for MeanFlow and flow matching (Section 4.5). We see this as a natural next step and added it to the Outlook (Section 5).

5. Setting the anchor weight μ. The anchor keeps each optimized state close to the state the prior's transport has reached, so μ trades measurement fit against staying on the prior's trajectory. It helps most when the measurements are noisy or incomplete: on the pixel priors, tuning selects a positive anchor for denoising and super-resolution and none for the other tasks, and on the latent priors, where the decoder already regularizes, it selects no anchor. The normalization by the dimension dd keeps μ comparable across resolutions. Appendix H.2 sweeps μ, and the revision adds a comparison with time-dependent anchors, including FLOWER's schedule, which do not improve on a constant one. A practical default is μ = 0, raised for strongly noisy measurements.

6. Blind inverse problems. Yes, in principle, when the unknown operator belongs to a differentiable parameterized family. In blind deblurring, for example, the blur kernel itself is unknown. Eq. (13) can then be extended to optimize the kernel jointly with the generative state at each stage, with a prior or regularizer on the kernel, as BlindDPS (Chung et al., 2023) does with diffusion priors. A MeanFlow prior keeps the advantage of differentiating through a single terminal prediction in this setting. We have not evaluated this extension and see it as a natural direction for future work.

We hope these answers address the reviewer's questions, and we thank the reviewer again for the helpful review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We thank all three reviewers for their constructive reviews. They led to the following changes in the revised paper; each reviewer's points are answered in the individual responses. Section numbers refer to the revised paper.

  • Equal-cost comparison (Section 4.4, Appendix K; WhU5, rSAb, yJUX). With every other method given RHSO's wall-clock time on pMF-L/16, RHSO has the best five-task mean in every metric.
  • Time and peak memory for every method on all four priors (Appendix I; rSAb, yJUX).
  • Controlled 2-D study with known data, run on flow-matching and MeanFlow priors trained identically (Section 4.1, Appendix L; WhU5).
  • FLOWER as an additional baseline on both pixel-space priors (Section 4.3; WhU5). With it included, RHSO is best in 32 of 60 task–metric cells and wins 221 of 270 pairwise comparisons.
  • Time-dependent anchors, including FLOWER's schedule, compared with the constant anchor (Appendix H.2; WhU5).
  • Corollary 7: the late-stage result behind Section 3.3 (Proposition 4) holds for data arbitrarily close to a low-dimensional manifold (Appendix B; rSAb).
  • Writing: Section 4 restructured and its densest passages rewritten in shorter sentences (WhU5, yJUX); Appendix B restructured with an overview of the results, a plain-language statement before each result, and the longer proofs split into labelled steps (rSAb, yJUX).
  • Outlook: adds comparisons with other flow-map families as priors (yJUX).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.