acceptodds
Under review as a conference paper at ICLR 2027

RestoreSteer: Adaptive Flow Control for Faithful Image Restoration

Abstract

Flow-based image editors provide strong generative priors for image restoration, but perceptually plausible outputs can deviate from the scene supported by a degraded observation, limiting restoration fidelity. We regard a restoration as faithful when it preserves scene information supported by the observation without reproducing the degradation itself. The available evidence and safely editable quantities change along the sampling trajectory: early states permit changes to scene formation but expose only coarse evidence; intermediate states provide more reliable boundaries while geometry remains editable; and late states support precise residual measurements but tolerate only small corrections. We introduce RestoreSteer, a training-free controller that matches the correction target and magnitude to these sampling phases. RouteCal redirects early scene formation using observation-supported structure and coarse appearance; StructAlign corrects persistent geometric disagreement after reliable boundaries emerge; and ObsRefine optimizes a bounded late latent residual through the frozen decoder and forward degradation model. Each stage evaluates bounded candidates before modifying the sampling state. RestoreSteer requires neither backbone retraining nor a fixed pseudo-clean target obtained by inverting the observation. Across six restoration tasks on Qwen-Image-Edit and FLUX.2, RestoreSteer improves PSNR, SSIM, and LPIPS over native sampling in every evaluated task-backbone setting. Component ablations further show complementary contributions from the three sampling phases.

Primary area:
generative models
Keywords:
Image Restoration, Flow Models, Trajectory Control, Observation Consistency, Training-Free Inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3pTuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Diffusion and flow models can generate realistic images, but in image restoration they may produce content inconsistent with the original observations, such as altered object layouts, hallucinated details, or incorrectly reconstructed structures. This paper proposes RestoreSteer, a training-free method that applies different trajectory corrections at different stages of flow-based generation. Each correction first evaluates a candidate result before deciding whether to modify the actual generation trajectory. This combines the generative prior with input observation constraints without training or changing the original model parameters.

Critical strengths

The method is intuitively effective, and the experiments demonstrate improvements across a relatively broad range of tasks and backbones.

Critical weaknesses
  1. Repeated trajectory endpoint estimation slows generation.
  2. The experiments do not examine paired-seed comparisons or stability.
Questions for authors
  1. Could you rerun the key Native/FlowSteer/RestoreSteer comparisons using the same per-image seeds and hardware, and report confidence intervals for paired differences or results from a few repeated seeds? In particular, please verify the differences involving Qwen’s native performance and the cases where the FLUX baseline degrades substantially.
  2. Were the conditional latent-space adaptations of PnP-Flow/FlowSteer first validated on tasks for which these methods are applicable and the operators are accurate? Were the different methods given reasonable tuning opportunities under the same approximate degradation model? Please provide details of the adaptations and supporting validation, rather than only citing the original methods.
7 comments

We sincerely appreciate the time you have taken to review our work and your constructive feedback. Please find our responses and additional results below.

W1. Repeated trajectory endpoint estimation slows generation.

Response to W1.

We appreciate your careful consideration of inference efficiency and acknowledge that RestoreSteer introduces additional computation. The overhead comes from repeated VAE decoding and backward passes during Adam-based latent refinement, as well as additional backbone evaluations at selected intervention steps. We recognize this additional inference cost as a practical limitation of the current method, despite its training-free design.

W2. The experiments do not examine paired-seed comparisons or stability.

Response to W2.

We are grateful for your attention to the reliability of the comparisons. We acknowledge that the earlier and later experiments used inconsistent seed conventions: Qwen Native was evaluated with seed 12345 for every image, while later experiments adopted a predetermined image-order-based rule and reused the earlier Native results without aligning seeds. Neither convention was chosen through seed or hyperparameter tuning to obtain more favorable results.

To address this inconsistency, we have completed a common-seed re-evaluation using seed 12345 for every image and method. RestoreSteer's improvements over Native remain across all six tasks and both backbones. Q1 presents the matched experimental conditions, complete results, image-paired confidence intervals, and an additional FLUX.2 comparison at seed 54321.

Q1. Could you rerun the key Native/FlowSteer/RestoreSteer comparisons using the same per-image seeds
and hardware, and report confidence intervals for paired differences or results from a few repeated
seeds? In particular, please verify the differences involving Qwen’s native performance and the
cases where the FLUX baseline degrades substantially.

Response to Q1.

Thank you for your careful consideration of our work and for your constructive comments and suggestions. We have re-evaluated the key methods with a common per-image seed and report the complete results together with paired confidence intervals.

Comparison conditions. Within each backbone, every method uses the same 50 input images per task, per-image seed 12345, restoration instruction, output size, and evaluation protocol. All experiments use NVIDIA A800 80GB PCIe GPUs with fixed software and model configurations. Qwen uses 40 sampling steps and FLUX.2 uses 50; PnP-Flow retains its own 100-iteration, five-sample solver budget.

Complete absolute results. Tables R1–R2 show the full comparison, with task columns split across two tables for readability. Bold marks the best displayed value within each backbone and task, including displayed ties.

Table R1. Fixed-seed denoising, dehazing, and low-light results.

BackboneMethodDenoising PSNR ↑Denoising SSIM ↑Denoising LPIPS ↓Dehazing PSNR ↑Dehazing SSIM ↑Dehazing LPIPS ↓Low light PSNR ↑Low light SSIM ↑Low light LPIPS ↓
QwenNative15.150.3350.40813.060.3600.47513.540.4460.314
Qwen+ PnP-Flow23.750.5720.63717.670.7250.32715.140.6320.278
Qwen+ FlowSteer25.470.7110.26217.010.7370.19713.890.6800.282
Qwen+ Ours24.640.6870.33919.400.7700.18017.500.6840.272
FLUX.2Native17.510.6320.23214.030.6020.33215.190.7080.250
FLUX.2+ PnP-Flow20.450.4750.77618.060.6890.42617.560.6530.330
FLUX.2+ FlowSteer20.830.5030.42016.550.7000.19214.620.6530.304
FLUX.2+ Ours25.800.7830.17419.800.8030.13518.400.7840.189

Table R2. Fixed-seed SR, deblurring, and colorization results.

BackboneMethodSR PSNR ↑SR SSIM ↑SR LPIPS ↓Deblurring PSNR ↑Deblurring SSIM ↑Deblurring LPIPS ↓Colorization PSNR ↑Colorization SSIM ↑Colorization LPIPS ↓
QwenNative17.190.4520.35618.390.4820.28410.580.2610.760
Qwen+ PnP-Flow25.680.6910.39613.530.3600.76122.560.7000.619
Qwen+ FlowSteer19.390.5430.64719.230.6250.38316.860.4710.671
Qwen+ Ours24.890.6790.28624.520.7310.25519.360.6310.397
FLUX.2Native15.130.3890.42018.080.4970.36118.100.6690.286
FLUX.2+ PnP-Flow20.900.5020.49518.450.5500.62713.910.4770.785
FLUX.2+ FlowSteer12.150.1051.14215.730.2510.88611.770.2580.691
FLUX.2+ Ours21.130.5490.35021.100.5760.31224.610.8870.164

Additional seed on FLUX.2. We repeated Native, FlowSteer, and RestoreSteer at seed 54321 on the same 50 images per task, with other settings unchanged. The repeat retains RestoreSteer's improvements over Native on all six FLUX.2 tasks (Table R3) and is reported separately from the seed-12345 confidence intervals.

Table R3. Additional FLUX.2 comparison at seed 54321.

TaskMethodPSNR ↑SSIM ↑LPIPS ↓
DenoisingNative17.180.6480.234
DenoisingFlowSteer20.830.5030.420
DenoisingRestoreSteer25.700.7790.187
HazeNative14.200.6100.343
HazeFlowSteer16.550.7000.192
HazeRestoreSteer20.570.8330.119
Low lightNative12.620.6280.313
Low lightFlowSteer14.620.6530.304
Low lightRestoreSteer16.090.7130.244
SRNative14.420.3600.449
SRFlowSteer12.110.1051.145
SRRestoreSteer20.810.5230.355
BlurNative17.980.5120.350
BlurFlowSteer15.870.2600.860
BlurRestoreSteer21.150.5850.308
ColorizationNative17.400.6300.299
ColorizationFlowSteer11.650.2610.692
ColorizationRestoreSteer23.480.8620.170

(Continued from Reply 2/7)

Image-paired confidence intervals. Tables R4–R6 compare methods with Native on six tasks and two backbones. We bootstrap 50 image pairs 10,000 times, reporting pointwise 95% CIs (2.5th/97.5th percentiles). Generation and resampling seeds are 12345. Positive PSNR/SSIM and negative LPIPS differences favor the listed method.

Table R4. RestoreSteer minus Native.

BackboneTaskΔPSNR (dB), 95% CIΔSSIM, 95% CIΔLPIPS, 95% CI
QwenDenoising+9.49 [+8.71, +10.31]+0.3520 [+0.3164, +0.3874]-0.0685 [-0.1102, -0.0285]
QwenHaze+6.33 [+5.28, +7.39]+0.4093 [+0.3777, +0.4415]-0.2951 [-0.3420, -0.2477]
QwenLow light+3.96 [+3.01, +4.93]+0.2379 [+0.1966, +0.2782]-0.0423 [-0.0679, -0.0187]
QwenSR+7.70 [+6.82, +8.64]+0.2274 [+0.1893, +0.2658]-0.0698 [-0.1036, -0.0390]
QwenBlur+6.12 [+5.33, +6.91]+0.2488 [+0.2167, +0.2822]-0.0293 [-0.0454, -0.0135]
QwenColorization+8.77 [+7.64, +9.96]+0.3696 [+0.3189, +0.4195]-0.3627 [-0.4214, -0.3003]
FLUX.2Denoising+8.29 [+7.31, +9.35]+0.1516 [+0.1340, +0.1694]-0.0573 [-0.0694, -0.0443]
FLUX.2Haze+5.77 [+4.76, +6.80]+0.2009 [+0.1800, +0.2220]-0.1965 [-0.2190, -0.1743]
FLUX.2Low light+3.21 [+2.45, +3.96]+0.0764 [+0.0588, +0.0942]-0.0610 [-0.0751, -0.0472]
FLUX.2SR+5.99 [+5.49, +6.52]+0.1606 [+0.1403, +0.1819]-0.0700 [-0.0880, -0.0507]
FLUX.2Blur+3.02 [+2.74, +3.31]+0.0784 [+0.0708, +0.0855]-0.0497 [-0.0579, -0.0413]
FLUX.2Colorization+6.51 [+5.59, +7.50]+0.2183 [+0.1960, +0.2421]-0.1217 [-0.1446, -0.1021]

Table R5. FlowSteer minus Native.

BackboneTaskΔPSNR (dB), 95% CIΔSSIM, 95% CIΔLPIPS, 95% CI
QwenDenoising+10.32 [+9.44, +11.25]+0.3761 [+0.3263, +0.4240]-0.1456 [-0.1962, -0.0961]
QwenHaze+3.94 [+2.77, +5.13]+0.3766 [+0.3348, +0.4187]-0.2783 [-0.3264, -0.2300]
QwenLow light+0.35 [-0.94, +1.60]+0.2340 [+0.1856, +0.2801]-0.0325 [-0.0595, -0.0057]
QwenSR+2.20 [+1.21, +3.19]+0.0910 [+0.0484, +0.1341]+0.2906 [+0.2460, +0.3327]
QwenBlur+0.84 [+0.31, +1.36]+0.1430 [+0.1193, +0.1658]+0.0984 [+0.0806, +0.1169]
QwenColorization+6.28 [+5.47, +7.12]+0.2099 [+0.1639, +0.2578]-0.0884 [-0.1225, -0.0554]
FLUX.2Denoising+3.32 [+2.61, +4.06]-0.1283 [-0.1698, -0.0895]+0.1881 [+0.1515, +0.2269]
FLUX.2Haze+2.52 [+1.43, +3.61]+0.0983 [+0.0687, +0.1273]-0.1397 [-0.1659, -0.1127]
FLUX.2Low light-0.57 [-1.49, +0.35]-0.0548 [-0.0780, -0.0290]+0.0546 [+0.0287, +0.0779]
FLUX.2SR-2.98 [-3.82, -2.16]-0.2834 [-0.3338, -0.2331]+0.7217 [+0.6844, +0.7581]
FLUX.2Blur-2.35 [-2.81, -1.90]-0.2465 [-0.2756, -0.2190]+0.5249 [+0.4853, +0.5628]
FLUX.2Colorization-6.33 [-7.34, -5.31]-0.4111 [-0.4504, -0.3695]+0.4050 [+0.3767, +0.4318]

Table R6. PnP-Flow minus Native.

BackboneTaskΔPSNR (dB), 95% CIΔSSIM, 95% CIΔLPIPS, 95% CI
QwenDenoising+8.60 [+7.73, +9.52]+0.2367 [+0.2047, +0.2699]+0.2294 [+0.1707, +0.2851]
QwenHaze+4.60 [+3.41, +5.82]+0.3648 [+0.3273, +0.4037]-0.1480 [-0.2132, -0.0818]
QwenLow light+1.60 [+0.34, +2.83]+0.1861 [+0.1410, +0.2323]-0.0363 [-0.0676, -0.0061]
QwenSR+8.50 [+7.59, +9.47]+0.2397 [+0.2008, +0.2789]+0.0396 [-0.0046, +0.0812]
QwenBlur-4.86 [-5.54, -4.21]-0.1219 [-0.1491, -0.0948]+0.4769 [+0.4546, +0.4995]
QwenColorization+11.98 [+11.19, +12.76]+0.4388 [+0.4129, +0.4643]-0.1404 [-0.1721, -0.1090]
FLUX.2Denoising+2.94 [+1.93, +4.01]-0.1569 [-0.1997, -0.1140]+0.5446 [+0.4964, +0.5919]
FLUX.2Haze+4.03 [+2.96, +5.08]+0.0874 [+0.0371, +0.1378]+0.0947 [+0.0069, +0.1853]
FLUX.2Low light+2.37 [+0.92, +3.90]-0.0547 [-0.1106, +0.0015]+0.0807 [+0.0301, +0.1297]
FLUX.2SR+5.77 [+5.16, +6.37]+0.1133 [+0.0875, +0.1388]+0.0743 [+0.0471, +0.1017]
FLUX.2Blur+0.37 [-0.14, +0.86]+0.0530 [+0.0214, +0.0840]+0.2652 [+0.2401, +0.2904]
FLUX.2Colorization-4.19 [-5.70, -2.60]-0.1921 [-0.2479, -0.1370]+0.4990 [+0.4505, +0.5471]
Reviewer_DnQpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces RestoreSteer, a training-free controller designed to improve the fidelity of image restoration using frozen flow-based generative models. Because generative priors often hallucinate details or deviate from the actual scene supported by a degraded observation, RestoreSteer treats restoration as an adaptive observation control problem. It applies targeted corrections that match the varying editability of the image along the sampling trajectory.

Critical strengths

Training-Free Versatility: The method requires neither backbone retraining nor task-specific restoration adapters, making it highly transferable across different flow editors and degradation types.

Interesting perspective on phase-dependent controllability: The conceptual separation into early scene formation, intermediate structural correction, and late observation consistency is intuitive and potentially generalizable beyond image restoration. It recognizes that corrections which may be effective early in the trajectory could be disruptive at later times.

Validated Architecture: Component ablations confirm that the three distinct phases (early calibration, middle alignment, and late refinement) provide complementary contributions to the final restoration.

Critical weaknesses

Limited evidence supporting the central hypothesis of phase-dependent control: Although the paper's main conceptual contribution is that different corrections should be applied at different stages of the flow trajectory, the experiments do not directly establish that the proposed phase assignment is necessary or optimal. The experiments demonstrate that RestoreSteer works, but provide weaker evidence that phase-adaptive control is responsible for its success.

The methodology appears highly heuristic and involves numerous design choices.: The overall approach combines spatial warping, task-specific appearance proposals, VAE encoding and decoding, latent residual optimization, multiple acceptance criteria, and manually designated intervention stages. While each design choice has some intuition, the main paper provides limited justification for why this particular combination is preferable to simpler alternatives. A systematic analysis is lacking and it looks more like a collection of heuristics that altogether work relatively well rather than a scientific approach to the problem.

The experimental comparisons are not entirely conclusive.: The authors compare against PnP-Flow and FlowSteer, adapting both methods to the conditional latent-space editors used in the experiments. However, these adaptations introduce an important question about whether the baseline implementations are equally well suited to the selected backbones and degradation settings. Some baseline results are surprisingly poor. For example, FlowSteer reduces FLUX.2 super-resolution PSNR from 15.01 dB under native sampling to 12.14 dB, while producing an LPIPS of 1.143. Such severe degradations raise questions about the effectiveness of the baseline adaptation.

Computational overhead is insufficiently characterized in the main paper: RestoreSteer requires repeated calls to the generative backbone, optimization of spatial transformations, multiple VAE encoding/decoding operations, and several gradient-based optimization steps. These operations can introduce substantial computational overhead relative to native sampling and simpler guidance approaches. The main paper does not quantify inference time, additional model evaluations, or memory requirements.

Questions for authors

The ablation experiments primarily evaluate the removal of individual components. Have you considered experiments that vary the timing of these interventions, swap their order, or apply a common correction strategy throughout sampling?

Could you provide a more fine-grained ablation study to identify which of the proposed operations (spatial warping, task-specific appearance adjustments, latent encoding differences, etc.) are essential?

What is the computational overhead relative to native sampling and the competing methods, in terms of both wall-clock time and model evaluations?

10 comments

We sincerely appreciate the time and care you have devoted to reviewing our work and your constructive feedback. We respond to each concern and question below.

W1. Limited evidence supporting the central hypothesis of phase-dependent control: Although the
paper's main conceptual contribution is that different corrections should be applied at different
stages of the flow trajectory, the experiments do not directly establish that the proposed phase
assignment is necessary or optimal. The experiments demonstrate that RestoreSteer works, but provide
weaker evidence that phase-adaptive control is responsible for its success.

Response to W1.

Thank you for your thoughtful assessment of phase-dependent control. We would respectfully clarify that RestoreSteer's contribution includes not only phase-specific coordination, but also the design of RouteCal, StructAlign, and ObsRefine. Table 2 evaluates these proposed mechanisms through component removal. We have additionally compared module orders and evaluated a common correction strategy at similar runtime. Q1 presents these comparisons.

W2. The methodology appears highly heuristic and involves numerous design choices.: The overall
approach combines spatial warping, task-specific appearance proposals, VAE encoding and decoding,
latent residual optimization, multiple acceptance criteria, and manually designated intervention
stages. While each design choice has some intuition, the main paper provides limited justification
for why this particular combination is preferable to simpler alternatives. A systematic analysis is
lacking and it looks more like a collection of heuristics that altogether work relatively well
rather than a scientific approach to the problem.

Response to W2.

We appreciate your careful comments on the methodological design and the need for a more systematic justification. We would like to clarify that the combination is guided by a common requirement: correcting observation-supported restoration errors without treating the degraded input as a clean target.

Spatial warping addresses geometric correspondence, while task-specific appearance proposals use the information available under each degradation. VAE encoding and decoding translate these image-space corrections into the frozen editor’s latent representation, and bounded residual optimization refines the correction locally. Observation consistency alone does not determine a faithful restoration, so the structure and detail criteria check properties that a lower observation error alone cannot capture. The phase-specific design further matches these corrections to early scene formation, intermediate geometric adjustment, and late observation refinement.

To assess the benefit of this design, we compare RestoreSteer with simpler alternatives. At comparable runtime, RestoreSteer achieves higher PSNR and SSIM than longer native sampling and uniformly spaced ObsRefine-only guidance on all three tested tasks (Q1, Table R5). Q1 also exchanges the module order while keeping all three correction mechanisms active. Q2 reports comparisons of spatial warping, appearance proposals, and proposal refinement. Together, these comparisons provide a more systematic basis for the combination, beyond the intuition behind its individual steps.

W3. The experimental comparisons are not entirely conclusive.: The authors compare against PnP-Flow
and FlowSteer, adapting both methods to the conditional latent-space editors used in the
experiments. However, these adaptations introduce an important question about whether the baseline
implementations are equally well suited to the selected backbones and degradation settings. Some
baseline results are surprisingly poor. For example, FlowSteer reduces FLUX.2 super-resolution PSNR
from 15.01 dB under native sampling to 12.14 dB, while producing an LPIPS of 1.143. Such severe
degradations raise questions about the effectiveness of the baseline adaptation.

Response to W3.

We are grateful for your attention to baseline reliability. We have checked the update calculations and editor interfaces, tested both adaptations with known operators, and examined a backbone-specific FlowSteer schedule adjustment. The seed-aligned real-image scores appear in Table R3; the numbers quoted in your review refer to the original comparison.

Adaptations and implementation checks. FlowSteer's official implementation uses FLUX.1-dev, a text-to-image model, with inversion and cached attention for image conditioning. Qwen-Image-Edit and FLUX.2 instead receive the degraded image and instruction through native editing interfaces, replacing that conditioning route. We retain FlowSteer's decode–correct–encode update, task-specific correction rules and strengths, and relative-step schedules.

PnP-Flow retains data fitting, fresh-noise interpolation, time-dependent denoising, and noise averaging, with 100 iterations and five noise samples per iteration. Its state uses the editor's clean latent coordinates; data-fitting gradients pass through the frozen decoder. Both editors remain frozen and use the same observation and instruction.

Controlled-input tests check correction residuals, scheduled steps and strengths, latent write-back, and PnP-Flow's denoising formula. FLUX.2 checks additionally cover latent packing, VAE conversion, reference-image tokens, and guidance-velocity reconstruction. Disabling FlowSteer's correction gives pixel-identical final images in all 200 paired comparisons—50 SR and 50 blur inputs per backbone—checking that integration preserves native sampling when no correction is applied. Appendix C.2 describes both adaptations; their reproduction code was included in the original supplementary material.

Validation with known operators. To check end-to-end restoration when the forward operator exactly matches input construction, we generated inputs from the benchmark's 50 clean references per task: ×4 area downsampling for SR and an 11 × 11 Gaussian blur with σ = 2 and circular boundaries for deblurring. Both baselines receive the same observations and matching operators. Clean references are used only for offline construction and evaluation.

Table R1. Known-operator validation, 50 images per group.

BackboneTaskMethodPSNR ↑SSIM ↑LPIPS ↓
QwenSRNative19.100.5220.263
QwenSRPnP-Flow29.130.8220.185
QwenSRFlowSteer20.310.6650.428
QwenBlurNative17.160.4370.306
QwenBlurPnP-Flow13.740.3880.746
QwenBlurFlowSteer23.400.8690.226
FLUX.2SRNative15.320.4070.368
FLUX.2SRPnP-Flow21.240.5320.441
FLUX.2SRFlowSteer12.270.1131.121
FLUX.2BlurNative20.010.6330.323
FLUX.2BlurPnP-Flow19.020.5860.605
FLUX.2BlurFlowSteer16.770.3160.888

These tests reproduce the key real-image trends: PnP-Flow improves SR PSNR/SSIM on both backbones; FlowSteer improves Qwen PSNR/SSIM on both tasks, but on FLUX.2 it remains below Native in all three metrics on both tested tasks. They provide controlled-task evidence alongside the numerical and interface checks.

(Continued from Reply 2/5)

FlowSteer on Qwen and FLUX.2. FlowSteer improves mean PSNR and SSIM over Native on all six Qwen tasks, whereas its benefits on FLUX.2 are less consistent. To clarify this difference, Table R2 compares the intervention positions and native scheduler noise levels under the same reference step ratios.

We retain the reference schedules, including the published SR and blur windows in Table 2 of FlowSteer; haze and low light use the task extensions described in Appendix C.2. Each row lists the intervention start, strength-change, and end positions in that order. Steps are one-based, with actual progress τ = (s − 1)/(N − 1); σ is averaged over 50 images per task.

Table R2. FlowSteer intervention positions and scheduler noise levels across backbones.

TaskReference ratios (%)Qwen steps (N = 40)Qwen mean σFLUX.2 steps (N = 50)FLUX.2 mean σ
Denoising50/75/9521/30/380.594/0.344/0.06125/38/480.840/0.624/0.179
Haze50/70/8521/28/340.599/0.410/0.21425/35/430.845/0.701/0.471
Low light50/70/8521/28/340.633/0.445/0.23825/35/430.883/0.764/0.551
SR50/70/8521/28/340.631/0.443/0.23725/35/430.880/0.759/0.545
Blur70/80/9028/32/360.440/0.308/0.15435/40/450.755/0.642/0.444
Colorization40/50/9517/21/380.692/0.599/0.06221/25/480.883/0.846/0.186

The same reference ratios correspond to higher native scheduler σ on FLUX.2 across all six tasks. FlowSteer corrects the image decoded from the current noisy latent, while we evaluate and correct the predicted clean endpoint. Its correction therefore operates under different noise conditions on the two backbones. For SR, for example, it compares a downsampled, still-noisy prediction with the observed LR image; the update can respond to intermediate noise as well as restoration errors, introducing artifacts into subsequent generation. This helps explain its less consistent FLUX.2 results and agrees with FlowSteer's noise-sensitivity analysis and discussion of empirical schedules.

To examine a targeted schedule adjustment, we tested Qwen-reference σ windows on all six FLUX.2 tasks. We map Qwen's three intervention noise coefficients to the closest positions on FLUX.2's native sampling grid. Only the intervention positions change; the model, native scheduler, observation operators, correction rule, and strengths remain fixed. All configurations use the same 50 images per task and seed 12345. Table R3 reports the effect of this adjustment.

Table R3. Original and adjusted FlowSteer windows on FLUX.2.

TaskMethod / configurationFlowSteer steps: start / change / endPSNR ↑SSIM ↑LPIPS ↓
DenoisingNative—17.510.6320.232
DenoisingFlowSteer: original ratios25 / 38 / 4820.830.5030.420
DenoisingFlowSteer: Qwen-reference σ39 / 45 / 4921.820.5160.427
DenoisingRestoreSteer—25.800.7830.174
HazeNative—14.030.6020.332
HazeFlowSteer: original ratios25 / 35 / 4316.550.7000.192
HazeFlowSteer: Qwen-reference σ39–40 / 44–45 / 4817.240.7810.155
HazeRestoreSteer—19.800.8030.135
Low lightNative—15.190.7080.250
Low lightFlowSteer: original ratios25 / 35 / 4314.620.6530.304
Low lightFlowSteer: Qwen-reference σ41 / 45 / 4815.290.7020.283
Low lightRestoreSteer—18.400.7840.189
SRNative—15.130.3890.420
SRFlowSteer: original ratios25 / 35 / 4312.150.1051.142
SRFlowSteer: Qwen-reference σ40–41 / 45 / 4818.110.2870.890
SRRestoreSteer—21.130.5490.350
BlurNative—18.080.4970.361
BlurFlowSteer: original ratios35 / 40 / 4515.730.2510.886
BlurFlowSteer: Qwen-reference σ45 / 47 / 4919.280.4630.498
BlurRestoreSteer—21.100.5760.312
ColorizationNative—18.100.6690.286
ColorizationFlowSteer: original ratios21 / 25 / 4811.770.2580.691
ColorizationFlowSteer: Qwen-reference σ35 / 39 / 4918.620.5170.533
ColorizationRestoreSteer—24.610.8870.164

The adjusted windows improve FlowSteer's mean PSNR and SSIM on all six tasks and reduce LPIPS on five, supporting intervention-noise conditions as a factor in the cross-backbone results. We will retain the original reference configuration in the main comparison and report both configurations and the complete tuning results in the appendix, with a main-text reference.

Reviewer_RwZzOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RestoreSteer, a training-free controller for flow-based image restoration. It introduces three stages: RouteCal for early structural correction, StructAlign for geometric alignment, and ObsRefine for late observation-consistency refinement. Experiments on six tasks with Qwen-Image-Edit and FLUX.2 show consistent improvements over native sampling.

Critical strengths

1.Table 1 shows improvements in PSNR, SSIM, and LPIPS across all six tasks and both backbones without retraining.

  1. Reasonable design and ablations. Phase-specific correction is intuitive, and Table 2 demonstrates the effectiveness of StructAlign and ObsRefine.
Critical weaknesses
  1. Insufficient evidence for phase-adaptive control. I think the novelty mainly lies in combining existing registration, observation guidance, and latent optimization techniques at different sampling stages. However, Table 2 only removes individual components without comparing alternative schedules or simpler guidance strategies under matched compute. Moreover, removing RouteCal improves low-light PSNR and LPIPS, questioning the general benefit of the proposed phase division.

  2. Baseline fairness and computational overhead. Appendix C.5 indicates different seed settings between native Qwen sampling and RestoreSteer. Also, adapted baselines such as FlowSteer perform surprisingly poorly on FLUX.2. I think these comparisons require stronger justification. Furthermore, RestoreSteer approximately doubles inference time, making its advantage over simpler methods less convincing without compute-matched comparisons.

Questions for authors
  1. Can the authors compare phase-adaptive control against fixed or uniformly spaced observation guidance under the same computational budget?

  2. Why are different seeds used for native Qwen and RestoreSteer? How were the adapted FlowSteer and PnP-Flow implementations validated?

  3. Why does removing RouteCal improve low-light restoration in Table 2? Under what conditions is early calibration beneficial?

7 comments
AuthorsOpenReview ↗

Replay 1/6

We sincerely appreciate the time you have taken to review our work and your constructive feedback. We respond to your concerns and questions below, following the order of your review.

W1. I think the novelty mainly lies in combining existing registration, observation guidance, and
latent optimization techniques at different sampling stages. However, Table 2 only removes
individual components without comparing alternative schedules or simpler guidance strategies under
matched compute. Moreover, removing RouteCal improves low-light PSNR and LPIPS, questioning the
general benefit of the proposed phase division.

Response to W1.

We are grateful for your thoughtful comments on our methodological contribution. Phase-specific coordination is indeed an important aspect of RestoreSteer. We would respectfully clarify, however, that our contribution extends beyond the timing of interventions: the design of RouteCal, StructAlign, and ObsRefine is itself part of the methodological novelty.

Contribution and component ablations. Specifically, RouteCal jointly calibrates early scene layout and coarse appearance; StructAlign introduces feedback-based geometric correction; and ObsRefine performs bounded late-stage latent refinement for observation consistency. Table 2 therefore evaluates these proposed correction mechanisms themselves by removing each component in turn.

Additional schedules and simpler controls. We appreciate your suggestion to examine alternative schedules alongside the component ablations. Q1 (Tables R2–R4) compares pairwise module-order swaps, uniformly spaced versus late-only ObsRefine, and longer native sampling or uniform guidance at similar runtime. These comparisons retain the corresponding correction rules and optimization limits.

Low-light restoration. Thank you for drawing attention to this result. We agree that the role of early calibration deserves a clearer explanation. RouteCal is designed to address early scene drift; on the evaluated low-light images, native Qwen sampling already preserves the scene layout reasonably well. In Q3, we discuss how this limited structural drift relates to the observed result and clarify the conditions under which early calibration is beneficial.

AuthorsOpenReview ↗

Replay 2/6

W2. Appendix C.5 indicates different seed settings between native Qwen sampling and RestoreSteer.
Also, adapted baselines such as FlowSteer perform surprisingly poorly on FLUX.2. I think these
comparisons require stronger justification. Furthermore, RestoreSteer approximately doubles
inference time, making its advantage over simpler methods less convincing without compute-matched
comparisons.

Response to W2.

We value your careful assessment of baseline fairness and inference cost. We have addressed these concerns through the seed-aligned comparison, baseline checks, and runtime comparisons below.

Seed settings. The earlier Qwen Native runs and the later experiments used different preset seed conventions, and we acknowledge that these should have been consistent. We have now repeated Native, PnP-Flow, FlowSteer, and RestoreSteer with seed 12345 for every evaluation image. RestoreSteer's improvements over Native remain across both backbones and all six tasks. For completeness, we explain the experiment history and present the full comparison in Q2.

FlowSteer adaptation and FLUX.2 results. We retain FlowSteer's observation-correction rule and task-dependent schedule, with decode–correct–encode updates at corresponding relative sampling positions. The observation is supplied through each editor's native image-conditioning interface; Q2 explains this adaptation and its validation.

Across the six FLUX.2 tasks, FlowSteer improves dehazing and improves denoising PSNR with a perceptual-quality trade-off, but performs less well on the other four tasks relative to Native. The same pattern remains after seed alignment.

Noise levels at matched intervention ratios. We retain the intervention ratios from FlowSteer's published task-specific schedules (Table 2 of FlowSteer): 70%–90% for blur and 50%–85% for SR. Table R1 compares scheduler noise levels at these same ratios on both backbones. The step columns show the actual correction positions recorded in the FlowSteer execution traces as s/N, with steps counted from 1. Relative progress is τ = (s − 1)/(N − 1), where s is the one-based step and N is 40 for Qwen or 50 for FLUX.2. These published ratios are mapped to actual steps using the implementation's rounding rule. For each image, σ is obtained from the native scheduler at that intervention step; the table reports means over 50 images per task.

Table R1. FlowSteer's published intervention ratios, executed sampling steps, and corresponding scheduler noise levels.

TaskPublished intervention ratioQwen stepQwen actual progressQwen σFLUX.2 stepFLUX.2 actual progressFLUX.2 σ
SR50%21/4051.28%0.63125/5048.98%0.880
SR70%28/4069.23%0.44335/5069.39%0.759
SR85%34/4084.62%0.23743/5085.71%0.545
Blur70%28/4069.23%0.44035/5069.39%0.755
Blur80%32/4079.49%0.30840/5079.59%0.642
Blur90%36/4089.74%0.15445/5089.80%0.444

Table R1 uses blur and SR as illustrative cases: the same published intervention ratios correspond to higher scheduler noise levels σ on FLUX.2 than on Qwen. FlowSteer applies observation correction to the image decoded from the current noisy latent, while our RestoreSteer evaluate and correct the predicted clean endpoint. At these higher-noise stages, forcing the degraded version of a still-noisy prediction to match the observed low-quality input can introduce artifacts and disrupt subsequent generation (Sections 4.1–4.2 of FlowSteer). This illustrates a broader mismatch between sampling-step ratios and noise levels across backbones, providing a plausible explanation for FlowSteer's weaker performance across multiple FLUX.2 restoration tasks.

Inference cost. RestoreSteer's additional time is spent on alignment, local latent refinement, and correction checks. To help assess this cost, we compare it with longer native sampling and repeated observation guidance at similar runtime. Table R2 shows higher PSNR and SSIM for RestoreSteer on all three tested tasks.

Q1. Can the authors compare phase-adaptive control against fixed or uniformly spaced observation
guidance under the same computational budget?

Response to Q1.

Thank you for your careful consideration of our work and for your constructive comments and suggestions. On Qwen-Image-Edit, we compare simpler controls at similar runtime, ObsRefine timing, and module-order swaps.

Comparable runtime. We compare the full method with two simpler uses of a similar inference budget. Native-80 doubles the native 40 sampling steps without adding restoration control. Uniform-12 applies ObsRefine alone at uniformly spaced steps, without RouteCal or StructAlign. It retains the observation loss, correction checks, and optimization limits, with the correction-strength sequence interpolated to twelve interventions.

For the runtime comparison, all variants use the same 50 images per task, prompts, output sizes, seed 12345, and GPU model. Measured time includes sampling, correction, and image encoding/decoding; model loading and metric evaluation are excluded. Bold indicates the best displayed value within each task and, where applicable, within each backbone/task in subsequent tables.

RestoreSteer's quality scores in Table R2 follow the fixed-seed main comparison in Tables R5–R6, using the matched suite's seed-12345 Full outputs for Qwen SR and haze. Inference times were measured separately under the runtime-comparison protocol.

Table R2. Restoration quality and measured inference cost.

TaskMethodPSNR↑SSIM↑LPIPS↓s/image↓
HazeNative-4013.060.3600.47552.28
HazeNative-8013.150.3600.468104.10
HazeUniform-1214.560.4030.456102.26
HazeRestoreSteer19.400.7700.180105.94
LowlightNative-4013.540.4460.31490.86
LowlightNative-8013.580.4570.313180.57
LowlightUniform-1216.280.5680.253209.70
LowlightRestoreSteer17.500.6840.272199.58
SRNative-4017.190.4520.35686.70
SRNative-8017.090.4480.364172.46
SRUniform-1218.920.4900.342188.62
SRRestoreSteer24.890.6790.286181.45

Table R2 directly addresses the computational-budget question: at similar inference time, RestoreSteer achieves higher PSNR and SSIM than either longer native sampling or uniform observation guidance on all three tasks. The benefit therefore reflects the correction strategy, rather than additional sampling time alone.

Timing with an unchanged correction rule. Uniform-8 and Late-8 both apply ObsRefine alone, with neither RouteCal nor StructAlign. To examine timing specifically, we keep the observation loss, ordered correction strengths, correction checks, eight scheduled correction steps, and optimization limits unchanged. Uniform-8 spreads these corrections over sampling, while Late-8 places them in the final part of the run.

Table R3. ObsRefine-only timing comparison.

TaskObsRefine-only schedulePSNR ↑SSIM ↑LPIPS ↓s/image ↓
HazeUniform-814.210.3920.45585.77
HazeLate-814.640.4060.45583.60
Low lightUniform-815.470.5390.266172.61
Low lightLate-815.800.5480.266172.02
SRUniform-818.310.4770.347154.51
SRLate-818.810.4830.345151.17

Late placement improves PSNR and SSIM on all three tested tasks without increasing runtime, while LPIPS remains close.

Swapping the module order. We additionally exchange the module assignments across the early (steps 5–14), middle (15–24), and late (25–39) windows on Qwen SR and haze. All three modules remain active, retaining their correction rules, strength sequences, and optimization limits.

Table R4. Pairwise module-order swaps on Qwen SR and haze.

TaskModule orderPSNR ↑SSIM ↑LPIPS ↓
SRStructAlign → RouteCal → ObsRefine24.610.6720.295
SRRouteCal → ObsRefine → StructAlign24.840.6540.323
SRObsRefine → StructAlign → RouteCal24.760.6510.343
SRRouteCal → StructAlign → ObsRefine (original)24.890.6790.286
HazeStructAlign → RouteCal → ObsRefine18.080.6920.243
HazeRouteCal → ObsRefine → StructAlign18.600.7490.185
HazeObsRefine → StructAlign → RouteCal19.120.7530.189
HazeRouteCal → StructAlign → ObsRefine (original)19.400.7700.180

The tested module-order swaps in Table R4 can reduce restoration quality, with particularly clear losses when StructAlign precedes RouteCal on haze. Together with the ObsRefine-only timing comparison, these results support matching correction types to sampling phases.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Changes in the Revised Manuscript

We sincerely thank all three reviewers for the time and care you have devoted to reviewing our work and for your constructive feedback. Your comments have helped us clarify the methodological contribution, improve comparison fairness, and strengthen the experimental support for our design. We summarize the main revisions below. All section, table, and appendix references refer to the revised manuscript.

  1. Clarified the second contribution in Section 1. We revised its wording to make clear that the design of RouteCal, StructAlign, and ObsRefine is part of our methodological contribution, alongside their phase-adaptive coordination.

  2. Standardized the seed protocol and updated the main comparison (Section 5.2, Table 1; Appendix C.6). We adopted a common per-image generation seed across the compared methods, updated the quantitative results, and revised the accompanying analysis to address concerns about comparison fairness.

  3. Updated the component-ablation reference and interpretation (Section 5.4, Table 2). We updated the Full reference and clarified the complementary, task-dependent contributions of the three components, explicitly acknowledging the low-light exception instead of implying that every component benefits every task.

  4. Added module-order comparisons (Section 5.4, Table 3). We compared pairwise order swaps with all three components active to examine how assigning different corrections to sampling phases affects restoration.

  5. Added operation-level ablations (Section 5.4, Table 4). We separately examined spatial warping, appearance proposals, and proposal refinement to provide a more detailed account of their roles within the controller.

  6. Added comparisons at comparable measured runtime (Section 5.4, Table 6). We compared RestoreSteer with longer native sampling and uniformly spaced ObsRefine-only guidance to assess whether these simpler uses of additional computation can reproduce its restoration gains.

  7. Expanded the computational-cost analysis (Section 5.5, Table 5; Appendix C.5, Tables 12–14). We added a main-text account of runtime and peak memory, expanded the appendix with backbone-evaluation counts and memory measurements, and clarified the principal sources of inference overhead.

  8. Clarified baseline adaptation and added matched-operator validation (Appendix C.2, Table 9). We explained how the conditional editors' native image conditioning replaces FlowSteer's original inversion and cached-attention conditioning route, and added SR and deblurring validation for both baselines with known, matched forward operators.

  9. Added an explanation of FlowSteer's cross-backbone behavior (Section 5.2; Appendix C.3, Tables 10–11). We compared scheduler noise levels at the reference intervention positions and evaluated adjusted windows, while distinguishing the original reference settings from the Qwen-reference noise-matched configuration.

AuthorsOpenReview ↗

Summary of Rebuttal and Reviewer Discussions 1/2

We sincerely thank all three reviewers for their time, constructive feedback, and follow-up discussions. We have carefully responded to their concerns and questions, conducted additional experiments, and incorporated the relevant results and clarifications into the revised manuscript.

Below, we summarize each reviewer's initial concerns, our responses, and the subsequent feedback. Table numbers refer to the corresponding reviewer's rebuttal thread.

Reviewer RwZz — Rating: 2 → 3 (Weak Reject → Weak Accept)

Reviewer RwZz acknowledged the improvements across six tasks and two backbones without retraining, and described the phase-specific correction design as intuitive. The reviewer raised concerns about phase-adaptive control, baseline fairness, computational overhead, and the low-light ablation.

  • Phase-adaptive control (W1 & Q1): We compared uniform and late ObsRefine interventions (Table R3), evaluated alternative module orders (Table R4), and compared simpler correction strategies at comparable runtime (Table R2).
  • Seed consistency (W2 & Q2): We repeated Native, PnP-Flow, FlowSteer, and RestoreSteer using the same per-image seed (12345), reporting complete results across six tasks and two backbones (Tables R5–R6).
  • Baseline reliability (W2 & Q2): We clarified the adaptations of FlowSteer and PnP-Flow, conducted known-operator validation (Table R7), and analyzed FlowSteer's intervention noise levels across backbones (Table R1).
  • Computational overhead (W2 & Q1): We compared RestoreSteer with longer native sampling and uniform observation guidance at similar runtime (Table R2), obtaining higher PSNR and SSIM on all three tested tasks.
  • Low-light ablation (W1 & Q3): We explained the observed improvement when removing RouteCal, discussing the limited structural drift under native sampling and the limitations of using an input-derived exposure reference rather than a ground-truth exposure target.

Reviewer RwZz — Follow-up: The reviewer acknowledged that the additional experiments resolved most concerns regarding baseline fairness and computational overhead. The rating was subsequently increased from 2 to 3.

We sincerely appreciate the reviewer's reconsideration and encouraging response.

Reviewer DnQp — Rating: 2 → 3 (Weak Reject → Weak Accept)

Reviewer DnQp identified "Training-Free Versatility," "Interesting perspective on phase-dependent controllability," and "Validated Architecture" as strengths. The reviewer also raised four critical weaknesses concerning phase-dependent control, methodological design, baseline comparisons, and computational overhead, along with three technical questions.

  • Phase-dependent control (W1 & Q1): We evaluated alternative module orders (Table R4) and compared RestoreSteer with longer native sampling and uniformly spaced observation guidance at comparable runtime (Table R5).
  • Methodological design (W2): We clarified the roles of spatial alignment, appearance proposals, latent correction, and candidate verification, together with the motivation for applying different corrections at different sampling phases.
  • Operation-level ablations (Q2): We separately evaluated spatial warping, appearance proposals, and proposal refinement on SR and dehazing (Table R6).
  • Baseline reliability (W3): We provided implementation checks and known-operator validation (Table R1), analyzed FlowSteer's cross-backbone intervention noise levels (Table R2), and evaluated adjusted FLUX.2 intervention windows (Table R3).
  • Computational overhead (W4 & Q3): We reported wall-clock inference time (Table R7), backbone evaluations (Table R8), and peak GPU memory (Table R9), while explaining the main sources of additional computation.

Reviewer DnQp — Follow-up: The reviewer acknowledged that the detailed responses and extensive ablation studies addressed an important part of the concerns. The rating was subsequently increased from 2 to 3.

We sincerely appreciate the reviewer's follow-up and consideration of our additional experiments.

AuthorsOpenReview ↗

Summary of Rebuttal and Reviewer Discussions 2/2

Reviewer 3pTu — Rating: 3 (Weak Accept, Unchanged)

Reviewer 3pTu stated that "The method is intuitively effective, and the experiments demonstrate improvements across a relatively broad range of tasks and backbones." The reviewer raised concerns about inference speed, paired-seed comparisons, experimental stability, and baseline adaptation.

  • Seed consistency (W2 & Q1): We repeated the main comparisons with matched per-image seeds, inputs, and evaluation settings (Tables R1–R2).
  • Experimental stability (W2 & Q1): We conducted an additional FLUX.2 evaluation using seed 54321 (Table R3) and reported image-paired 95% bootstrap confidence intervals (Tables R4–R6). RestoreSteer's intervals relative to Native exclude zero in the favorable direction across all six tasks and both backbones.
  • Baseline adaptation and validation (Q2): We clarified the adaptations of FlowSteer and PnP-Flow and conducted validation using known degradation operators (Table R7), supported by implementation checks and reproduction details.
  • FlowSteer intervention schedules (Q1 & Q2): We analyzed scheduler noise levels across backbones (Table R8) and evaluated adjusted intervention windows on FLUX.2 (Table R9).
  • Inference efficiency (W1): We acknowledged the additional inference cost and explained its sources. Quantitative measurements are provided in our response to Reviewer DnQp (Tables R7–R9).

Reviewer 3pTu — Follow-up: The reviewer acknowledged that the additional experiments addressed some of the concerns and confirmed maintaining the rating at 3.

We sincerely appreciate the reviewer's follow-up and continued consideration.

Overall Summary

We have provided responses to the concerns and questions raised by all three reviewers, supported by additional experiments and corresponding manuscript revisions.

Following the rebuttal, two reviewers increased their ratings from 2 to 3, while the third maintained a rating of 3. We are grateful that all three reviewers revisited our responses and shared their follow-up assessments.

We sincerely thank everyone involved in the review process for their time, effort, and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.