- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes a framework of gradient-related optimization methods to solve image inverse problem based on pre-trained flow matching models. Specifically, the authors introduce a novel regularization term through a distance function and formulated the related optimization problem. The authors demonstrate that any gradient-based methods can be applied to solve the optimization problem and provide several algorithm examples including FlowDPS, FlowTSGD, FlowFTSGD, FlowBCD, and FlowADMM. These algorithms are evaluated on four image inverse problem tasks and three datasets, with FlowDPS, FlowBCD, and FlowADMM performing well.
Critical strengths
-
Interpreting the pre-trained flow matching model as the gradient of an explicit regularization term is meaningful and works as effective tools to involve any gradient-based methods to solve inverse problems.
-
The main experiments are reasonable, covering four kinds of inverse problems (super-resolution via bicubic and avgpool downsamplings, Gaussian and Motion deblurring) under three datasets (DIV2K, FFHQ, AFHQ) and four common image metrics (PSNR, SSIM, LPIPS, FID).
Critical weaknesses
-
Although the paper lists existing relevant methods in Related Works, it does not clearly separate its contributions into a new regularizer in the field of solving inverse problems or an optimization interpretation of an existing flow prior. I could not get the point about the paper's contribution or improvement especially on what existing methods could not do.
-
The comparison to baselines and ablation study are insufficient. The paper only compared two methods (FlowChef and FlowDPS). Some other related works like Flower [1] should be also added as baselines. Also, the ablation study for hyperparameter shows the trend that smaller leads to better performance in most metrics (PSNR, SSIM, FID). It remains unclear whether a much smaller would leads to better performace.
[1] Pourya et al. "Flower: A flow-matching solver for inverse problems", ICLR 2026.
Questions for authors
I list my questions below and if I misunderstood any part of the method, please correct me.
-
I'm confused about the definition of the velocity field in the manuscript: seems to have opposite signs in Theorem 3.1 and Definition 4.1. In Theorem 3.1 and Appendix A, the training loss leads to the velocity . However, in Definition 4.1 which is based on the proposition 1 in FlowDPS, the Tweedie's formula used the common velocity field , which leads to contradiction.
-
When deriving FlowFTSGD, the paper seems unifying two methods (FlowDPS and PnP-Flow), but it remains unclear that under what exact values of , FlowFTSGD would correspond to these two methods.
-
The inequality used in Line 632 to prove Theorem 3.1 is Jensen's Inequality, yes? It would be better to have clarification after Line 645.
-
There seems a typo in Line 695 (Equation (4.4), Line 4): Eqn. (4.4) is a single-line optimization problem and contains no "Line 4". Please correct it.
-
There seems a missing related work: Webber et al. "Solving Inverse Problems with Flow-based Models via Model Predictive Control", ICML 2026.
-
The paper describes four optimization methods with the iterative scheme but no pseudo-code algorithms. Adding some pseudo-code algorithms could improve clarity and reproducibility.
5 comments
I thank the authors for the detailed rebuttal. While some of my concerns have been addressed, several issues remain:
-
The response to W1 reiterates the goal of connecting generative models with optimization and interpreting the velocity field as a regularization gradient. However, it still fails to clarify the paper's novelty/improvement on what existing methods could not do and show what limitations it addresses.
-
W2 is also only partially addressed, since my original questions concerned two parts: insufficient baseline comparisons and ablation study. The response provides the ablation study results of hyperparameter but misses the additional comparisons with closely related methods such as Flower [1].
-
The revised version of Theorem 3.1 seems reasonable, but without an updated proof, its correctness could not be verified. Could the authors provide the complete revised derivation?
-
As for Q3-Q5, I could not locate the revisions and please let me know if I may have missed them in the current version.
As for Q6, I thank the authors for providing these pseudo-code algorithms which improves the paper's readability. Overall, I appreciate the authors’ clarifications and additional results, but the concerns above remain unresolved. Therefore, I maintain my original rating.
[1] Pourya et al. "Flower: A flow-matching solver for inverse problems", ICLR 2026.
[Q1] Our key insight is that the velocity field of a flow matching model can be interpreted as the gradient of an explicit regularization term , i.e., the generative prior enters the optimization as a well-defined regularizer rather than as sampling-time guidance.
-
It lets us embed a flow matching prior into classical optimization algorithms (BCD and ADMM, giving FlowBCD and FlowADMM), which come with established convergence analysis — something guidance-based solvers lack.
-
Because the prior is an explicit term in the objective, the regularization strength has a clear interpretation and can be set by standard parameter-selection principles rather than those guidance schedules.
-
Most researchers in the image processing community are more familiar with optimization than with probability and stochastic processes, so our paper serves as a bridge between generative modeling and this community.
[Q2] The following presents experimental comparisons on Flower, where the number of sampling steps is set to 28 for fairness, matching the other methods.
Avgpool SR
| Dataset | PSNR$\uparrow$ | SSIM$\uparrow$ | FID$\downarrow$ | LPIPS$\downarrow$ |
|---|---|---|---|---|
| AFHQ | 16.02 | 0.320 | 99.42 | 0.320 |
| FFHQ | 15.93 | 0.363 | 195.96 | 0.345 |
| DIV2K | 14.79 | 0.230 | 160.98 | 0.407 |
Bicubic SR
| Dataset | PSNR$\uparrow$ | SSIM$\uparrow$ | FID$\downarrow$ | LPIPS$\downarrow$ |
|---|---|---|---|---|
| AFHQ | 15.89 | 0.363 | 99.90 | 0.382 |
| FFHQ | 15.73 | 0.423 | 191.50 | 0.347 |
| DIV2K | 14.95 | 0.271 | 159.82 | 0.403 |
Gaussian Deblurring
| Dataset | PSNR$\uparrow$ | SSIM$\uparrow$ | FID$\downarrow$ | LPIPS$\downarrow$ |
|---|---|---|---|---|
| AFHQ | 16.31 | 0.278 | 59.08 | 0.205 |
| FFHQ | 15.90 | 0.320 | 141.71 | 0.188 |
| DIV2K | 15.46 | 0.238 | 117.76 | 0.223 |
Motion Deblurring
| Dataset | PSNR$\uparrow$ | SSIM$\uparrow$ | FID$\downarrow$ | LPIPS$\downarrow$ |
|---|---|---|---|---|
| AFHQ | 19.70 | 0.437 | 0.156 | |
| FFHQ | 18.89 | 0.429 | 62.02 | 0.145 |
| DIV2K | 18.33 | 0.401 | 53.91 | 0.142 |
[Q3] We have updated the revision, thank you for your time and efforts reviewing our paper.
Response to Reviewer t5uF
[W1] We aim to establish a connection between generative models and optimization and understand flow matching-based imaging inverse problems from an optimization perspective.
In classical optimization methods for imaging inverse problems, we consider minuGt(u)=λtF(u)+Rt(u), where F(u) is the data-fidelity objective and Rt(u) is the regularization term.
When solving it with gradient-based methods, we need ∇uGt(u)=λt∇uF(u)+∇uRt(u). In Section 3, we set Rt(u) as a distance-based function show that ∇uRt(u)=−vθ, where vθ is the pretrained velocity field in the unconditional flow matching model.
Now, both ∇uGt(u) and ∇uRt(u) are determined and we can apply gradient-based optimization methods directly.
[W2] To ensure good LPIPS performance, we set ρ0=40 in this paper. Numerical results for ρ0=25,30 show no performance improvement.
[Q1] We revised Theorem 3.1 as follows:
Let M be a data manifold. For any ϵ∼N(0,I) and u∈M, define z=atu+btϵ, where at,bt≥0. Let vθ(z,t) be the pre-trained neural network to estimate the velocity field in the flow matching model (see the overview in Appendix A). Set ℓt(z)=μtfM(ctz,σt)−2ηt∣∣z∣∣2. When the neural network is optimal trained--that is, when ∫01EϵEu∣∣a˙tu+b˙tϵ−vθ(atu+btϵ,t)∣∣2dt=0, it follows that ∇zℓt(z)=−vθ(z,t), where ct=at1,σt=atbt,μt=btat(a˙tbt−atb˙t),ηt=at2μt+btb˙t. In practice, ∫01EϵEu∣∣a˙tu+b˙tϵ−vθ(atu+btϵ,t)∣∣2dt>0 is a small constant after training, vθ(zt,t) approximates the negative gradients of the function ℓt(z) with respect to z.
[Q2] τt=bt for the FlowDPS, while PnP-Flow corresponds to the FlowTSGD.
[Q3-5] Yes, we used Jensen's Inequality to prove Theorem 3.1, which is added in revision. For Q4, we modify to Equation (4.3), Line 3. For Q5, we add this reference to related works. Thank you for the comments.
[Q6] Pseudocode of FlowTSGD
Input: at,bt,ζt,t∈[0,1], predicted velocity vθ(z,t), num,
t=0, δ=num1,
Sample z∼N(0,I),
For(t<1)
u←z−bt∇zℓt(z)=z+btvθ(z,t),
u←u−ζt∇uF(u),
t←t+δ,
z←Qt(u,z),
Output: z.
Pseudocode of FlowFTSGD
Input: at,bt,ζt,τt,t∈[0,1], predicted velocity vθ(z,t), num,
t=0, δ=num1,
Sample z∼N(0,I),
For(t<1)
u←z−bt∇zℓt(z)=z+btvθ(z,t),
v←u−ζt∇uF(u),
u←τtu+(1−τt)v,
t←t+δ,
z←Qt(u,z),
Output: z.
Pseudocode of FlowBCD
Input: at,bt,ζt,τt,ρt,λt,t∈[0,1], predicted velocity vθ(z,t), num,
t=0, δ=num1,
Sample z∼N(0,I), w∼N(0,I),
For(t<1)
u←Pt(z),
v←Pt(w),
u←z+btvθ(z,t)−ρt(u−v),
s←v−ζt[λt∇vF(v)+ρt(v−u)],
v←τtv+(1−τt)s,
t←t+δ,
z←Qt(u,z),
w←Qt(v,w).
Output: w.
Pseudocode of FlowADMM
Input: at,bt,ζt,τt,ρt,λt,t∈[0,1], predicted velocity vθ(z,t), num,
t=0, δ=num1,
Sample z∼N(0,I), w∼N(0,I),
Λ=0,
For(t<1)
u←Pt(z),
v←Pt(w),
u←z+btvθ(z,t)−ρt(u−v)+Λ,
s←v−ζt[λt∇vF(vt)+ρt(v−u)]−ζtΛ,
v←τtv+(1−τt)s,
Λ←Λ+ρt(v−u),
t←t+δ,
z←Qt(u,z),
w←Qt(v,w).
Output: w.