- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper focuses on solving inverse problems with flow matching priors. In particular, the authors propose VEG, which replaces the data-gradient step on the clean image with a step preconditioned by the Jacobian of the denoiser (computed with a vector–Jacobian product), and then re-noises to keep running the solver. This is motivated by a first-order analysis showing that the measurement shifts the conditional mean along a covariance-shaped direction, and that for the linear flow-matching path this covariance is proportional to the Jacobian of the denoiser. Experiments on AFHQ-Cat, CelebA and ImageNet (JiT) cover nonlinear blur, HDR and several linear inverse problems.
Critical strengths
-
The paper connects the clean-image correction to a clear statistical quantity: the first-order change of the conditional mean is a covariance-shaped direction, which for the linear path can be computed with one VJP of the denoiser. Even though the identities are known, putting them together gives a simple and cheap method.
-
The experiments are extensive (several datasets, operators and pretrained models) and include negative results, although some comparisons are missing (see below). In particular, the comparison with ordinary gradient corrections is convincing: one VEG correction gets higher PSNR than ten ordinary corrections (+0.58 dB) while using 0.64× the runtime, so the gain comes from the direction and not just from more optimization
Critical weaknesses
-
I believe the main weakness of the paper is the limited novelty. The gradient in eq. (15) is exactly DPS in the flow matching language (also explored in OT-ODE [2]), and Propositions 1 and 2 are well-known results (e.g., Peng et al. 2024 and Boys et al. 2024, both cited in the paper, and [3]). Furthermore, the algorithm is essentially the same splitting scheme as PnP-Flow, DiffPIR [4] and FLOWER, as the side-by-side comparison shows ( is the denoiser, ):
PnP-Flow (Martin et al., 2025), at each :
- Data step on the clean image: .
- Re-noise: .
- Denoise: .
VEG, at each :
- Data step on the clean image: , with , , and from an Armijo line search on .
- Re-noise: .
- Denoise: .
Up to re-indexing the loop, the only differences are the direction of the data step ( instead of , i.e., a gradient step preconditioned by the denoiser Jacobian) and the step-size rule (line search instead of a fixed schedule). The norm matching is not the contribution either, since Table 7 shows that the raw and theory-scaled VJPs perform as well or better. So I think the method is better described as "PnP-Flow/FLOWER with a covariance-preconditioned data step", which is a much narrower claim than "endpoint geometry". From this view, the natural missing comparison is the exact solve in the covariance metric, , which for linear is the MMPS [3] update ; similarly, if is the right geometry, I'd expect the proximal term in to use and not be isotropic. Also, the norm matching throws away the scaling of Proposition 2, so the step size is set by the line search and not by the theory (without backtracking the method diverges, Table 7).
Regarding the writing, the paper spends a lot of time on results and definitions that are not fundamental to the key message, and I found it hard to read. For instance, "endpoint geometry" is just with a new name, and the algorithm is only given in the appendix (Algorithm 1), although it would directly show what changes with respect to PnP-Flow/FLOWER.
-
The second main weakness is the experimental part. First, Table 1 doesn't include PnP-Flow, which is as efficient as VEG; its gradient step only needs , and the authors already adapted DAPS, FLOWER and D-Flow to the nonlinear setting. Second, the results don't show a clear improvement over the baselines: on the nonlinear benchmark (Table 1), VEG doesn't get the best PSNR or LPIPS in any of the four dataset–task groups (e.g., D-Flow is 3.4 dB better on CelebA HDR), and on JiT (Table 16) FLOWER gets higher PSNR in 4 out of 5 tasks. The linear results (Table 2) are more favorable, but use only 10 images per dataset and task. Since the paper is framed as a quality–compute trade-off, the comparison should be against methods with a similar runtime (see Clarification 1), and not mostly against much slower ones. Finally, Table 3 is hard to understand from the main text: the angle-matched random direction is only defined in Appendix B.4, and mixing development and held-out "cohorts" makes the rows hard to compare.
Overall, I am not convinced by the numerical experiments, and I believe the contribution is narrow, which leans me towards rejection.
Questions for authors
- The gradient in (6) is w.r.t. ? or ?
Clarifications (answers could affect my assessment)
-
How does VEG compare with FlowDPS [1] or OT-ODE [2] on the main AFHQ-Cat/CelebA benchmark? FlowDPS is only compared on JiT (Appendix E), and since the paper is framed as a quality–compute trade-off, I think methods with a similar runtime are the relevant comparison.
-
Can you compare against MMPS/TMPD-style [3] updates that actually solve with the Jacobian (e.g., a few CG steps using VJPs)? This would tell us whether the first-order approximation is the right trade-off.
-
Did you run JAPS on nonlinear blur and HDR? Both papers use these tasks, so a direct comparison seems feasible.
-
In Section 4.1 you state that, to first order, a noisy-state update (as in DPS) changes the clean image by (≈ ), while the clean-image correction gives (≈ ). This is stated without a derivation or discussion. Furthermore, can you explain intuitively why should be preferred over ?
Also, the noisy-state baseline in Table 3 / Appendix B.5 changes the step size, the line search and the continuation at the same time, so the gain can't be attributed to where this update is applied. Can you provide an experiment where only the placement changes?
[1] Kim, J., Kim, B. S., & Ye, J. C. (2025, October). Flowdps: Flow-driven posterior sampling for inverse problems. ICCV 2025
[2] Pokle, Ashwini, Matthew J. Muckley, Ricky TQ Chen, and Brian Karrer. "Training-free linear image inverses via flows." TMLR
[3] Rozet, François, Gérôme Andry, François Lanusse, and Gilles Louppe. "Learning diffusion priors from observations by expectation maximization." NeurIPS 2024
[4] Zhu, Yuanzhi, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. "Denoising diffusion models for plug-and-play image restoration." CVPR 2023
8 comments
Clarification 3: JAPS on Nonlinear Blur and HDR
We directly evaluated JAPS [3] on both nonlinear blur and HDR saturation using the same frozen OT-flow prior, images, and observations as VEG. As shown in Clarification 1, JAPS achieves slightly higher mean PSNR and lower LPIPS but requires 2.67× the mean runtime. On nonlinear blur, VEG achieves comparable PSNR at substantially lower computational cost. The comparison retains JAPS's sampling procedure while adapting its denoiser interface to the shared flow predictor. Detailed results are reported in Table 1, with implementation details in Appendix D.1 and a methodological comparison in Appendix B.7 (Table 12).
Clarification 4: DPS Relation and Correction Placement
VEG and DPS use proportional VJP guidance directions under the shared flow formulation, but they are not identical correction algorithms. Under affine re-noising, VEG's clean-image correction induces a proportional next-state VJP update, while finite-step scaling and acceptance can produce different corrections and trajectories. The previously discussed term instead describes same-time predictor reevaluation:
=-\alpha JJ^\top g+o(\alpha).$$ **This is not the actual DPS next-state update.** We therefore removed Proposition 3 and revised the DPS interpretation in Section 4.1. **Direct clean-image correction substantially reduces computation, with no statistically resolved PSNR difference in the matched placement comparison.** To isolate correction placement, we compare direct correction $c_n+\alpha d_n$ with state reevaluation $D_{t_n}(x_n+\alpha d_n)$, keeping the direction construction, local objective, Armijo policy, fallback, and noise-mixing continuation with shared Gaussian noise draws fixed. | Correction placement | PSNR ↑ | LPIPS ↓ | Time (s) ↓ | |---|---:|---:|---:| | Direct clean correction | 25.379 | 0.1627 | **15.83** | | Same-time state reevaluation | 25.450 | **0.1560** | 62.97 | On 40 development cases, the PSNR difference is unresolved, while state reevaluation achieves lower LPIPS at **3.98× runtime** because each line-search candidate requires predictor reevaluation. Thus, the advantage of direct correction is computational efficiency rather than intrinsically better reconstruction quality. We clarified this comparison in Section 6.3 and report the controlled results in Appendix B.5 (Table 10). **The same-sampler DPS comparison further shows that proportional VJP directions may not yield identical reconstruction performance.** VEG improves PSNR by **2.362 dB over DPS and 1.552 dB over DPS+Armijo**, with lower LPIPS and measurement NRMSE. These gains are consistent with differences in correction scaling and acceptance rather than a distinct VJP direction. We added these comparisons to Table 5 and detailed the baseline implementations in Appendix D.1. # W1: Novelty and Meaning of Endpoint Geometry **Our contribution is to characterize how measurements change the conditional mean of clean endpoints in a flow model and validate the resulting correction direction, rather than to introduce a new VJP or splitting procedure.** Endpoint geometry refers to the local variation of clean endpoints conditioned on the flow state, $q_t(x_1\mid x_t)$, characterized by its covariance $C_t$. Incorporating measurement $y$ reweights this distribution toward $p(x_1\mid x_t,y)$. Section 3 characterizes the resulting first-order mean change as $m_t'(0)=-\operatorname{Cov}_{q_t}(X_1,\ell_y(X_1))\approx-C_t\bar g_t$, which is proportional to the clean-prediction VJP under the ideal independent-Gaussian flow. While the underlying statistical and Jacobian identities are known, **we connect them to measurement-guided endpoint correction and investigate the reconstruction information carried by this direction.** **Tables 3 and 4 examine whether the flow-derived direction provides useful information beyond ordinary measurement descent and how that information affects reconstruction.** Table 3 tests whether its benefit comes from the direction itself, rather than correction magnitude, alignment with ordinary descent, or additional refinement. Table 4 investigates whether the direction can guide reconstruction in measurement-insensitive regions, showing improvements in missing or saturated regions where the measurement gradient vanishes, even when alternative corrections yield identical measurement predictions. **We substantially revised Section 6.3** to clarify the motivation, controlled conditions, and interpretation of both experiments, with further details in Appendices B and C.Cont.
W1: Norm Matching, Theoretical Scale, and Proximal Metric
The theoretical covariance interpretation motivates the correction direction, while norm matching and line search control its finite-step implementation. The factor relates the ideal conditional covariance to the clean-prediction Jacobian but does not prescribe the optimal step size for a learned predictor. Raw, theory-scaled, and norm-matched VJPs achieve similar PSNR with backtracking, whereas removing backtracking can cause divergence. Thus, norm matching is a practical safeguard rather than an independent source of reconstruction gains.
Likewise, the isotropic proximal penalty controls the correction magnitude rather than approximating the covariance-weighted metric . We clarified these roles and the independent-Gaussian assumptions in Sections 3.3–4.2 and Appendix A, with the scaling ablation reported in Appendix B.4 (Table 8).
Writing and Presentation Improvements
We substantially revised the manuscript to clarify the central idea and improve the presentation of the experimental evidence. Specifically, we clarified endpoint geometry in Sections 2–3, moved Algorithm 1 into Section 4.2, and substantially revised Section 6.3 to explain the motivation and controlled conditions of Tables 3–4, including the angle-matched random control and the distinction between development and held-out evaluations.
[1] Kim, J., Kim, B. S., and Ye, J. C. FlowDPS: Flow-Driven Posterior Sampling for Inverse Problems. ICCV, 2025.
[2] Martin, S., Gagneux, A., Hagemann, P., and Steidl, G. PnP-Flow: Plug-and-Play Image Restoration with Flow Matching. ICLR, 2025.
[3] Hen, L., Tirer, T., Giryes, R., and Abu-Hussein, S. Jacobian-Aware Posterior Sampling for Inverse Problems. TMLR, 2026.
[4] Rozet, F., Andry, G., Lanusse, F., and Louppe, G. Learning Diffusion Priors from Observations by Expectation Maximization. NeurIPS, 2024.
[5] Saad, Y., and Schultz, M. H. GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems. SIAM Journal on Scientific and Statistical Computing, 1986.
Thanks,
Authors
Dear Reviewer A36s,
Thank you for your detailed review. We have expanded baseline comparisons, clarified the theoretical interpretation, and improved the presentation of controlled experiments. We address your questions first, followed by the broader concerns.
Q1: Gradient Definition
The gradient in Eq. (6) is taken with respect to the predicted clean image ct, not the flow state xt. Specifically, gt=∇ctℓy(ct), whereas Eq. (15) computes JDt(xt)⊤gt=∇xtℓy(Dt(xt)) by the chain rule. We clarified this in Sections 2.2 and 4.1.
Clarification 1: Comparison with FlowDPS and Similar-Compute Methods
VEG achieves higher reconstruction quality than FlowDPS and PnP-Flow on the main nonlinear benchmark, including a substantial improvement over PnP-Flow at comparable runtime. We evaluated FlowDPS [1], PnP-Flow [2], and JAPS [3] under the shared pretrained OT-flow prior on 120 held-out cases, averaged equally across four dataset–task groups.
VEG improves mean PSNR by 1.886 dB over FlowDPS-NL and 2.593 dB over PnP-Flow-NL, with lower LPIPS in both comparisons. FlowDPS-NL is faster, while PnP-Flow-NL has comparable runtime. JAPS and D-Flow achieve higher fidelity but require 2.67× and 19.5× the mean runtime, respectively. These results support a favorable quality–compute trade-off rather than uniform superiority. Tables 1–2 report the full results, with adaptations in Appendix D.1.
Clarification 2: Comparison with MMPS/TMPD-Style Jacobian Solves
The benefit of exploiting the flow-derived covariance structure depends on how it is used, rather than simply on the number of Jacobian computations. We examine this through iterative corrections motivated by MMPS [4]. We first evaluated Conjugate Gradient (CG) with one and three iterations using a clean-prediction Jacobian covariance proxy. On 60 paired linear cases, VEG achieves 29.699 dB PSNR with 99 predictor VJPs, compared with 29.488 dB using 198 VJPs and 29.195 dB using 396 VJPs for CG-1 and CG-3. However, the learned Jacobian does not guarantee a symmetric positive-definite system, limiting standard CG. We report details in Appendix B.8 (Table 13).
To address this limitation, we further compare against GMRES [5], which handles nonsymmetric linear systems. MMPS likewise replaces CG with GMRES in its diffusion benchmark because learned denoiser Jacobians produce nonsymmetric, non-definite systems. We adapt this measurement-space formulation to the frozen flow predictor. We construct a measurement-space system using the clean-prediction Jacobian as a covariance proxy, combined with the local measurement Jacobian and observation-noise covariance. Let ct=Dt(xt), Jt=JDt(xt), Ht=Jh(ct), χt=(1−t)2/t, and Γ=(σy2+10−6)I. We define
M_t(u)&=\chi_tJ_t^\top u,\\\\ S_t(q)&=\Gamma q+H_tM_t(H_t^\top q),\\\\ q_t^{(k)}&=\operatorname{GMRES}_k(S_t,y-h(c_t);0),\\\\ \delta_t^{(k)}&=M_t(H_t^\top q_t^{(k)}),\quad k\in\{1,3\}. \end{aligned}$$ **GMRES solves the measurement-space system approximately using the observation residual, then maps the solution to a clean-image correction through $H_t^\top$ and $M_t$.** Additional iterations can refine its direction through further Jacobian products. We match VEG's proposal norm and retain its descent test, Armijo search, and noise-mixing continuation. The results cover **120 paired nonlinear cases**, with runtime measured on the same RTX 5080 GPU. | Method | PSNR ↑ | LPIPS ↓ | Time (s) ↓ | Predictor VJPs ↓ | |---|---:|---:|---:|---:| | **VEG** | 26.037 | 0.16050 | **15.51** | **99** | | GMRES-1 | 26.009 | 0.15983 | 26.83 | 297 | | GMRES-3 | **27.396** | **0.12685** | 38.65 | 495 | GMRES-1 produces a correction collinear with the original VJP, up to sign, without resolved quality improvement. **GMRES-3 improves PSNR by 1.359 dB and reduces LPIPS by 0.03365**, but requires **2.49× runtime and five times the predictor VJPs**. **These results suggest that effective use of the Jacobian-based covariance proxy can improve reconstruction, whereas more Jacobian computations alone do not guarantee improvement.** VEG remains a lower-compute alternative. Neither comparison constitutes an exact posterior-covariance solve or a reproduction of native MMPS/TMPD. We added these results and details in Appendix B.9 (Table 14).