- 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
- Repeated trajectory endpoint estimation slows generation.
- The experiments do not examine paired-seed comparisons or stability.
Questions for authors
- 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.
- 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
Reply 2/7
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.
| Backbone | Method | Denoising PSNR ↑ | Denoising SSIM ↑ | Denoising LPIPS ↓ | Dehazing PSNR ↑ | Dehazing SSIM ↑ | Dehazing LPIPS ↓ | Low light PSNR ↑ | Low light SSIM ↑ | Low light LPIPS ↓ |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen | Native | 15.15 | 0.335 | 0.408 | 13.06 | 0.360 | 0.475 | 13.54 | 0.446 | 0.314 |
| Qwen | + PnP-Flow | 23.75 | 0.572 | 0.637 | 17.67 | 0.725 | 0.327 | 15.14 | 0.632 | 0.278 |
| Qwen | + FlowSteer | 25.47 | 0.711 | 0.262 | 17.01 | 0.737 | 0.197 | 13.89 | 0.680 | 0.282 |
| Qwen | + Ours | 24.64 | 0.687 | 0.339 | 19.40 | 0.770 | 0.180 | 17.50 | 0.684 | 0.272 |
| FLUX.2 | Native | 17.51 | 0.632 | 0.232 | 14.03 | 0.602 | 0.332 | 15.19 | 0.708 | 0.250 |
| FLUX.2 | + PnP-Flow | 20.45 | 0.475 | 0.776 | 18.06 | 0.689 | 0.426 | 17.56 | 0.653 | 0.330 |
| FLUX.2 | + FlowSteer | 20.83 | 0.503 | 0.420 | 16.55 | 0.700 | 0.192 | 14.62 | 0.653 | 0.304 |
| FLUX.2 | + Ours | 25.80 | 0.783 | 0.174 | 19.80 | 0.803 | 0.135 | 18.40 | 0.784 | 0.189 |
Table R2. Fixed-seed SR, deblurring, and colorization results.
| Backbone | Method | SR PSNR ↑ | SR SSIM ↑ | SR LPIPS ↓ | Deblurring PSNR ↑ | Deblurring SSIM ↑ | Deblurring LPIPS ↓ | Colorization PSNR ↑ | Colorization SSIM ↑ | Colorization LPIPS ↓ |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen | Native | 17.19 | 0.452 | 0.356 | 18.39 | 0.482 | 0.284 | 10.58 | 0.261 | 0.760 |
| Qwen | + PnP-Flow | 25.68 | 0.691 | 0.396 | 13.53 | 0.360 | 0.761 | 22.56 | 0.700 | 0.619 |
| Qwen | + FlowSteer | 19.39 | 0.543 | 0.647 | 19.23 | 0.625 | 0.383 | 16.86 | 0.471 | 0.671 |
| Qwen | + Ours | 24.89 | 0.679 | 0.286 | 24.52 | 0.731 | 0.255 | 19.36 | 0.631 | 0.397 |
| FLUX.2 | Native | 15.13 | 0.389 | 0.420 | 18.08 | 0.497 | 0.361 | 18.10 | 0.669 | 0.286 |
| FLUX.2 | + PnP-Flow | 20.90 | 0.502 | 0.495 | 18.45 | 0.550 | 0.627 | 13.91 | 0.477 | 0.785 |
| FLUX.2 | + FlowSteer | 12.15 | 0.105 | 1.142 | 15.73 | 0.251 | 0.886 | 11.77 | 0.258 | 0.691 |
| FLUX.2 | + Ours | 21.13 | 0.549 | 0.350 | 21.10 | 0.576 | 0.312 | 24.61 | 0.887 | 0.164 |
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.
| Task | Method | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|---|
| Denoising | Native | 17.18 | 0.648 | 0.234 |
| Denoising | FlowSteer | 20.83 | 0.503 | 0.420 |
| Denoising | RestoreSteer | 25.70 | 0.779 | 0.187 |
| Haze | Native | 14.20 | 0.610 | 0.343 |
| Haze | FlowSteer | 16.55 | 0.700 | 0.192 |
| Haze | RestoreSteer | 20.57 | 0.833 | 0.119 |
| Low light | Native | 12.62 | 0.628 | 0.313 |
| Low light | FlowSteer | 14.62 | 0.653 | 0.304 |
| Low light | RestoreSteer | 16.09 | 0.713 | 0.244 |
| SR | Native | 14.42 | 0.360 | 0.449 |
| SR | FlowSteer | 12.11 | 0.105 | 1.145 |
| SR | RestoreSteer | 20.81 | 0.523 | 0.355 |
| Blur | Native | 17.98 | 0.512 | 0.350 |
| Blur | FlowSteer | 15.87 | 0.260 | 0.860 |
| Blur | RestoreSteer | 21.15 | 0.585 | 0.308 |
| Colorization | Native | 17.40 | 0.630 | 0.299 |
| Colorization | FlowSteer | 11.65 | 0.261 | 0.692 |
| Colorization | RestoreSteer | 23.48 | 0.862 | 0.170 |
Reply 3/7
(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.
| Backbone | Task | ΔPSNR (dB), 95% CI | ΔSSIM, 95% CI | ΔLPIPS, 95% CI |
|---|---|---|---|---|
| Qwen | Denoising | +9.49 [+8.71, +10.31] | +0.3520 [+0.3164, +0.3874] | -0.0685 [-0.1102, -0.0285] |
| Qwen | Haze | +6.33 [+5.28, +7.39] | +0.4093 [+0.3777, +0.4415] | -0.2951 [-0.3420, -0.2477] |
| Qwen | Low light | +3.96 [+3.01, +4.93] | +0.2379 [+0.1966, +0.2782] | -0.0423 [-0.0679, -0.0187] |
| Qwen | SR | +7.70 [+6.82, +8.64] | +0.2274 [+0.1893, +0.2658] | -0.0698 [-0.1036, -0.0390] |
| Qwen | Blur | +6.12 [+5.33, +6.91] | +0.2488 [+0.2167, +0.2822] | -0.0293 [-0.0454, -0.0135] |
| Qwen | Colorization | +8.77 [+7.64, +9.96] | +0.3696 [+0.3189, +0.4195] | -0.3627 [-0.4214, -0.3003] |
| FLUX.2 | Denoising | +8.29 [+7.31, +9.35] | +0.1516 [+0.1340, +0.1694] | -0.0573 [-0.0694, -0.0443] |
| FLUX.2 | Haze | +5.77 [+4.76, +6.80] | +0.2009 [+0.1800, +0.2220] | -0.1965 [-0.2190, -0.1743] |
| FLUX.2 | Low light | +3.21 [+2.45, +3.96] | +0.0764 [+0.0588, +0.0942] | -0.0610 [-0.0751, -0.0472] |
| FLUX.2 | SR | +5.99 [+5.49, +6.52] | +0.1606 [+0.1403, +0.1819] | -0.0700 [-0.0880, -0.0507] |
| FLUX.2 | Blur | +3.02 [+2.74, +3.31] | +0.0784 [+0.0708, +0.0855] | -0.0497 [-0.0579, -0.0413] |
| FLUX.2 | Colorization | +6.51 [+5.59, +7.50] | +0.2183 [+0.1960, +0.2421] | -0.1217 [-0.1446, -0.1021] |
Table R5. FlowSteer minus Native.
| Backbone | Task | ΔPSNR (dB), 95% CI | ΔSSIM, 95% CI | ΔLPIPS, 95% CI |
|---|---|---|---|---|
| Qwen | Denoising | +10.32 [+9.44, +11.25] | +0.3761 [+0.3263, +0.4240] | -0.1456 [-0.1962, -0.0961] |
| Qwen | Haze | +3.94 [+2.77, +5.13] | +0.3766 [+0.3348, +0.4187] | -0.2783 [-0.3264, -0.2300] |
| Qwen | Low light | +0.35 [-0.94, +1.60] | +0.2340 [+0.1856, +0.2801] | -0.0325 [-0.0595, -0.0057] |
| Qwen | SR | +2.20 [+1.21, +3.19] | +0.0910 [+0.0484, +0.1341] | +0.2906 [+0.2460, +0.3327] |
| Qwen | Blur | +0.84 [+0.31, +1.36] | +0.1430 [+0.1193, +0.1658] | +0.0984 [+0.0806, +0.1169] |
| Qwen | Colorization | +6.28 [+5.47, +7.12] | +0.2099 [+0.1639, +0.2578] | -0.0884 [-0.1225, -0.0554] |
| FLUX.2 | Denoising | +3.32 [+2.61, +4.06] | -0.1283 [-0.1698, -0.0895] | +0.1881 [+0.1515, +0.2269] |
| FLUX.2 | Haze | +2.52 [+1.43, +3.61] | +0.0983 [+0.0687, +0.1273] | -0.1397 [-0.1659, -0.1127] |
| FLUX.2 | Low light | -0.57 [-1.49, +0.35] | -0.0548 [-0.0780, -0.0290] | +0.0546 [+0.0287, +0.0779] |
| FLUX.2 | SR | -2.98 [-3.82, -2.16] | -0.2834 [-0.3338, -0.2331] | +0.7217 [+0.6844, +0.7581] |
| FLUX.2 | Blur | -2.35 [-2.81, -1.90] | -0.2465 [-0.2756, -0.2190] | +0.5249 [+0.4853, +0.5628] |
| FLUX.2 | Colorization | -6.33 [-7.34, -5.31] | -0.4111 [-0.4504, -0.3695] | +0.4050 [+0.3767, +0.4318] |
Table R6. PnP-Flow minus Native.
| Backbone | Task | ΔPSNR (dB), 95% CI | ΔSSIM, 95% CI | ΔLPIPS, 95% CI |
|---|---|---|---|---|
| Qwen | Denoising | +8.60 [+7.73, +9.52] | +0.2367 [+0.2047, +0.2699] | +0.2294 [+0.1707, +0.2851] |
| Qwen | Haze | +4.60 [+3.41, +5.82] | +0.3648 [+0.3273, +0.4037] | -0.1480 [-0.2132, -0.0818] |
| Qwen | Low light | +1.60 [+0.34, +2.83] | +0.1861 [+0.1410, +0.2323] | -0.0363 [-0.0676, -0.0061] |
| Qwen | SR | +8.50 [+7.59, +9.47] | +0.2397 [+0.2008, +0.2789] | +0.0396 [-0.0046, +0.0812] |
| Qwen | Blur | -4.86 [-5.54, -4.21] | -0.1219 [-0.1491, -0.0948] | +0.4769 [+0.4546, +0.4995] |
| Qwen | Colorization | +11.98 [+11.19, +12.76] | +0.4388 [+0.4129, +0.4643] | -0.1404 [-0.1721, -0.1090] |
| FLUX.2 | Denoising | +2.94 [+1.93, +4.01] | -0.1569 [-0.1997, -0.1140] | +0.5446 [+0.4964, +0.5919] |
| FLUX.2 | Haze | +4.03 [+2.96, +5.08] | +0.0874 [+0.0371, +0.1378] | +0.0947 [+0.0069, +0.1853] |
| FLUX.2 | Low light | +2.37 [+0.92, +3.90] | -0.0547 [-0.1106, +0.0015] | +0.0807 [+0.0301, +0.1297] |
| FLUX.2 | SR | +5.77 [+5.16, +6.37] | +0.1133 [+0.0875, +0.1388] | +0.0743 [+0.0471, +0.1017] |
| FLUX.2 | Blur | +0.37 [-0.14, +0.86] | +0.0530 [+0.0214, +0.0840] | +0.2652 [+0.2401, +0.2904] |
| FLUX.2 | Colorization | -4.19 [-5.70, -2.60] | -0.1921 [-0.2479, -0.1370] | +0.4990 [+0.4505, +0.5471] |
Reply 1/7
We sincerely appreciate the time you have taken to review our work and your constructive feedback. Please find our responses and additional results below.
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.
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.