acceptodds
Under review as a conference paper at ICLR 2027

Global Before Local: Search and Selection in Flow-Based Inverse Problems

Abstract

When more inference computation is available, should an inverse solver refine one reconstruction longer or explore alternatives? We study this through global-before-local (GBL) inference, which advances multiple candidates, selects a subset, and completes the retained candidates with the same solver. Broader search can find adequate candidates yet still lose when selection discards them. We separate best-candidate advantage from selection loss, quantify the recovery needed to beat equal-work refinement, and use a Bayes benchmark to distinguish information limits from selector error. GBL with measurement-based selection gains 3.88 dB in mean peak signal-to-noise ratio (PSNR) over refinement in CelebA phase retrieval. Predictive selection improves reconstruction quality in held-out inpainting and exceeds refinement in AFHQ irregular masking. On CelebA 4× point-decimation super-resolution, development data show that even the best searched candidate trails the same-budget single trajectory and forecast a 1.30 dB mean-PSNR deficit for predictive selection; held-out evaluation confirms the ordering with a 1.15 dB deficit. These results show that the value of broader search depends not only on the candidates it finds, but also on how much of that advantage selection preserves.

Primary area:
generative models
Keywords:
inverse problems, flow matching, inference-time computation, candidate selection, image reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6u5EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses a question: when more inference computation is available, should an inverse solver refine one reconstruction longer or explore alternatives? Their answer is to design a global-before-local (GBL) inference, which advances multiple candidates, selects a subset, and completes the retained candidates with the same solver.

Critical strengths
  1. The question of whether a generative inverse solver should choose longer refinement or broader search is analyzed quite clearly. The paper demonstrates that the optimal strategy varies depending on the task, offering valuable insights for future work.
  2. The designs of method are reasonable. By optimizing the sampling strategy, this method achieves improvements over Flower and FlowDPS on some settings.
Critical weaknesses
  1. The authors may have overlooked the impact of guidance strength of generative inverse solvers on reconstruction performance. For a baseline, a hyperparameter such as FlowDPS and FLOWER can be used to adjust the strength of its data consistency, which has a significant impact on the results. In this paper, the number of steps in the baseline methods was adjusted to facilitate a fair comparison. However, maintaining the original default guidance strength under these modified settings may not be appropriate. Furthermore, the authors utilized a set of development images to select strategies for the test set and train GBL-P. In practice, methods such as FlowDPS and FLOWER could also leverage such development images to determine the optimal guidance strength for that specific dataset. Consequently, the above factors may make the comparison unfair.
  2. In a real-world setting, the limitations of this method are as follows: (1) The approach relies on development cohorts to determine the strategy applied to all images in the test set. However, in practical scenarios, the development cohorts may differ, and the optimal configuration for each individual image might also vary. (2) Most of the results in this paper focus on strategies optimized for PSNR. However, the balance between fidelity and perceptual is also an issue worth considering.
  3. The readability of the experimental section could be improved. Many of the settings and terms introduced (e.g., single predictor and predictor ensemble in Table 3) are unexplained. The figures and tables follow the logical flow of the explanation, but they lack a direct comparison between the baseline and the two proposed models across all settings.
Questions for authors
  1. Please provide a table directly comparing the baseline with the two proposed models across all settings (on at least one dataset).
  2. Please specify the guidance strength used for your methods and the baseline.
  3. Please provide a table comparing the performance of the baseline method (Flower or FlowDPS) and the two models proposed in this paper as the guidance strength varies.
1 comment

Dear Reviewer 6u5E,

Thank you for your detailed review and helpful suggestions. We have added direct comparisons, tested guidance sensitivity and independent tuning, and clarified the experimental settings.

W3+Q1. Direct Comparison and Experimental Clarity

GBL-P achieves higher mean PSNR than both GBL-M and the single-path baseline on AFHQ-Cat central and irregular inpainting under the same 548 solver-update budget. On central inpainting, the single path, GBL-M, and GBL-P achieve 13.87, 14.46, and 14.86 dB, respectively. On irregular inpainting, their PSNR values are 15.22, 15.11, and 15.80 dB. On phase retrieval, GBL-M has slightly higher PSNR than GBL-P. Table 4 provides the direct comparison across all three settings, including LPIPS. We also clarified the predictor settings in Table 3 and Appendix B.2: the single predictor uses 64 training images, while the ensemble averages five predictors, each trained on four of five image-level folds of 192 training images.

W1+Q2–Q3. Guidance Strength and Fairness

GBL-P's PSNR advantage over the single-path baseline on AFHQ-Cat central inpainting remains across the tested guidance strengths and after independent tuning. Our main comparisons use the same guidance settings for GBL and the single path: FlowDPS uses three corrections per update with η=15\eta=15, and FLOWER uses likelihood weight B=1B=1 with an Armijo line search, as specified in Section 5.1.

We further tested three FlowDPS guidance strengths on AFHQ-Cat central inpainting, generating new trajectories and training a separate GBL-P predictor for each strength.

Methodη=7.5\eta=7.5η=15\eta=15η=30\eta=30
Single10014.26314.35614.483
Single54813.84213.87013.904
GBL-M14.47514.46114.731
GBL-P14.74014.92814.951

PSNR (dB) on 128 held-out images. Note that Table 5 (in the revised manuscript) also reports LPIPS.

GBL-P improves mean PSNR over Single548 by 0.90–1.06 dB across these strengths. When GBL-P and Single548 select guidance independently using the same 64 development images, both choose η=30\eta=30 (FlowDPS guidance step size). For GBL-P, each development image is evaluated using a predictor trained without that image. GBL-P then achieves a 1.046 dB gain, improving PSNR on 97 of 128 images. The CelebA phase-retrieval gain also remains 3.680 dB after independent guidance selection. These results are detailed in Appendices D.6–D.7 and Table 21.

W2. Generalization and Perceptual Quality

We extended the evaluation to examine when broader search is useful and how PSNR and perceptual quality may differ. In CelebA central inpainting with independently selected guidance, GBL-P achieves the lowest LPIPS, while the single path retains the highest PSNR, as shown in Table 20. We also added a super-resolution experiment in Section 5.5, where development results correctly predict that single-path inference will outperform GBL-P on separate held-out images. Adapting the candidate count and selection time to individual images and changes in data distributions remains future work.

Thanks,

Authors

Reviewer_zK4nOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies global-before-local inference from various angles and designs various evaluation criteria to decide what helps and where.

Critical strengths

None of the results presented can be understood clearly. There are only scattered numbers here and there.

Critical weaknesses
  • [Presentation] The entire content of the paper seems to be AI-generated without human review, particularly in terms of choice of words, section titles, exposition, and flow.

  • Abstract itself if very dense and cryptic with a lot of unrelated numbers.

  • Main contributions are not stated (at all) and it looks very difficult to disentangle the content.

Questions for authors
  • Can you summarize the main punchline of the paper?
  • Can you state your contributions more clearly?
1 comment

Dear Reviewer zK4n,

Thank you for your comments. We have revised the paper to make the main idea, contributions, and results easier to understand.

Q1. Main Finding

We study whether it is better to spend more inference steps on one reconstruction or search across several candidates. GBL runs multiple trajectories to an intermediate step, selects a subset, and continues only the selected ones. Our main finding is that broader search can improve reconstruction, and that its benefit depends on both the quality of the candidates found and the ability to select among them. We provide an analysis that explains when broader search is useful, when better selection can help, and when a single longer trajectory is preferable. In our experiments, GBL-M improves mean PSNR by 3.88 dB in CelebA phase retrieval, GBL-P improves selection in inpainting, and our analysis correctly predicts the advantage of single-path inference in super-resolution.

Q2. Main Contributions

We measure the gains from search and losses from selection, determine how much selection must improve to beat a single path, and use a Bayes analysis to separate limits of the available information from those of the selection rule. We further introduce GBL-P, which predicts final reconstruction quality from partial trajectories, and use development results to predict the mean performance difference on held-out images.

We have rewritten the Abstract and Introduction to explain these contributions clearly, simplified the method and notation in Section 3, and organized Section 5 around the purpose of each experiment. Figures 2–3 show the main findings, and Table 4 directly compares GBL-M, GBL-P, and single-path inference.

Thanks,
Authors

Reviewer_GjMsOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Global before local (GBL) for exploring the trade-offs between exploration and exploitation in the context of test-time scaling (more specifically, inverse problem reconstructions) in the context of diffusion models. The key idea is to initialize a bunch of denoising trajectories and select a candidate that is further denoised. Several variants of this global-before-local strategy, like GBL-M and GBL-P, are presented, and the empirical results are explored in the context of inverse problems in diffusion models.

Critical strengths

I think this aspect of diffusion models is quite interesting to study, and the paper does a decent job of systematically exploring it. Some ideas, like completing the discarded candidates offline to tell the selected failures apart from search failures, are quite intuitive.

Critical weaknesses

My main concerns are as follows:

Presentation: The paper is simply hard to follow. I had to spend a considerable amount of time dealing with jargon and the empirical protocol. For instance, Section 3, which describes the global vs. local strategy, makes the text unnecessarily complicated by introducing unnecessary notations. This all could have been explained in simple text.

Novelty and Empirical Results:

– Can the authors elaborate on the core novel contributions of the paper? In part, this question results from the presentation of the paper. Moreover, this idea of initializing multiple trajectories and choosing the best has been explored in a number of prior works [1,2,3], which are not cited in the main text. The authors should position their work relative to this existing line of work.

– There is almost no comparison with Sequential Monte Carlo (SMC) and particle-based baselines in the paper. Comparisons are primarily made with some variants of the proposed method. Can the authors compare GBL against one SMC baseline at equal compute on the tasks presented in the paper?

– Moreover, how does the performance compare against general state-of-the-art methods for solving inverse problems like NDTM [4], DiffPIR [5], DAPS[6]?

References:

[1] Practical and Asymptotically Exact Conditional Sampling in Diffusion Models, Wu et al.

[2] Monte Carlo Guided Denoising Diffusion Models for Bayesian Linear Inverse Problems, Cardoso et al.

[3] A General Framework for Inference-time Scaling and Steering of Diffusion Models, Singhal et al.

[4] NDTM: “Variational Control for Guidance in Diffusion Models.” Pandey et al.

[5] DiffPIR: “Denoising Diffusion Models for Plug-and-Play Image Restoration.” Zhu et al.

[6] DAPS: “Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing.” Zhang et al.

Questions for authors

See weaknesses

1 comment

Dear Reviewer GjMs,

Thank you for your comments. We have clarified the paper's main contribution, simplified the presentation, and added particle-based comparisons across three inverse-problem settings.

Q1. Core Contributions and Prior Work

Our main question is when searching across multiple trajectories is more useful than spending the same update budget on a single trajectory, and how the quality of the candidates and the selection rule together determine the outcome. Our contributions are to separate the quality gained through search from the quality lost during selection, quantify how much selection must improve to beat the single-path baseline, and use a Bayes analysis to distinguish information limits from those of the selection rule. We complete discarded candidates for evaluation and use this analysis to explain different experimental outcomes and predict the mean PSNR difference between GBL-P and single-path inference on held-out images, as described in Section 5.5. We have also expanded Section 2 to position our work relative to TDS [1], MCGdiff [2], FK Steering [3], and related pruning methods.

Q2. Comparison with Particle-Based Methods

We added an FK-style particle-resampling baseline based on [3], using intermediate rewards, multinomial resampling, and independent continuation. We evaluated two particle allocations on three inverse-problem settings, keeping the observations, pretrained prior, and solver fixed within each task. The results on 128 images per setting are summarized below.

TaskMultistart 5×100FK 5×100FK 4×137GBL-M
CelebA phase28.535 / 0.105027.656 / 0.119427.344 / 0.131130.171 / 0.0794
CelebA central16.156 / 0.238216.234 / 0.234916.703 / 0.239916.611 / 0.2356
AFHQ central14.610 / 0.489614.518 / 0.495814.418 / 0.506914.461 / 0.4979
Solver updates500500548548
Model evaluations500500548555

Each entry: PSNR (dB) / LPIPS

On CelebA phase retrieval, GBL-M improves mean PSNR over FK 5×100 and FK 4×137 by 2.516 and 2.828 dB, respectively, with lower LPIPS in both comparisons. Both improvements remain positive under adjusted paired confidence intervals. On central inpainting, the differences between GBL-M and FK resampling are not statistically clear. FK 5×100 also shows no clear improvement over independent multistart across the three settings.

These results are included in Section 5.2 and Appendix C.7, with the full comparison reported in Table 15. All methods use at most 548 solver updates, and we report the additional model evaluations separately.

Q3. Comparison with Other Inverse Solvers

We reviewed the official implementations and pretrained model configurations of NDTM, DiffPIR, and DAPS. Their pretrained priors and resolutions do not directly match our CelebA-128 and AFHQ-Cat-256 settings. Changing the generative prior would introduce another variable into the comparison, so we have not added direct numerical results for these methods. Appendix C.1 reports additional comparisons with D-Flow, FlowDPS, and FLOWER in Table 9 under their stated settings and computational budgets, while our main experiments isolate the effect of computation allocation using a fixed prior and solver.

Thanks,
Authors

References

[1] Wu et al., "Practical and Asymptotically Exact Conditional Sampling in Diffusion Models," NeurIPS, 2023.

[2] Cardoso et al., "Monte Carlo Guided Denoising Diffusion Models for Bayesian Linear Inverse Problems," ICLR, 2024.

[3] Singhal et al., "A General Framework for Inference-time Scaling and Steering of Diffusion Models," ICML, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.