- 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
-
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.
-
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
- [Clarification] Equation (13) appears to be mathematically equivalent to MPC- under a simple reparameterisation:
In MPC-, let denote the optimised control and define the uncontrolled flow step as
Making the substitution
transforms the MPC- 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
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 on pMF-L/16, each with its own tuned . The first follows FLOWER's schedule: with FLOWER's , it uses , which increases with . The second uses its reciprocal, , 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.
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): 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+Δvθ(xtk,tk),
where xˉ does not depend on u (for a MeanFlow prior, xˉ=Tθ(xtk;tk→tk+1)). RHSO instead moves the state before the network transports it, xtk+1=Ek(xtk+δk), so its control is an input to the network rather than an increment added to its output. Substituting x′=xˉ+Δu turns the MPC-Δt stage objective into Eq. (13) at time tk+1, with μ=2d/(λmΔ2), but the correction still acts at a different place relative to the network. The case N=1 makes this explicit. MPC-Δt samples xˉ=x0+vθ(x0,0) and then solves minu∥u∥2+λΦ(xˉ+u), a regularized correction in image space that never backpropagates through the prior. RHSO solves minx0Φ(Tθ(x0;0→1))+2dμ∥x0−s0∥2 from the sampled noise s0, 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.