acceptodds
Under review as a conference paper at ICLR 2027

Formation-Consistent Flow Matching for Single-Image Defocus Deblurring

Abstract

Single-image defocus deblurring seeks to recover an all-in-focus image from spatially varying defocus blur. Recent generative restoration methods based on Flow Matching can produce visually plausible results but may introduce hallucinated details that are not well supported by the observed blur. To address this limitation, we propose Formation-Consistent Flow Matching (FCFlow), a training framework that incorporates an effective defocus formation model into Flow Matching endpoint optimization. FCFlow learns an effective spatially varying circle-of-confusion (CoC) representation directly from the blurred input without requiring depth or CoC annotations. At a randomly sampled flow time, the predicted velocity is used to obtain a sharp endpoint proposal through single-step extrapolation. The proposal is decoded into image space and reblurred using the estimated CoC map, producing a formation discrepancy that is converted into a bounded latent endpoint correction. This correction refines endpoint optimization, thereby shaping the learned velocity field toward restorations that better explain the observed blur. The formation-aware endpoint optimization branch is omitted at inference, retaining standard Flow Matching sampling without additional formation-related overhead. Extensive experiments on DPDD, LFDOF, and RealDOF demonstrate state-of-the-art performance with 20-step inference, while strong single-step results further highlight the effectiveness of FCFlow in the low number of function evaluations (NFE) regime.

Primary area:
generative models
Keywords:
Flow Matching, Defocus Deblurring, Circle-of-Confusion, Endpoint Feedback

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_vMQHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FCFlow, which introduces defocus formation feedback into latent-space flow matching from blurry to sharp images. The method first predicts effective circle-of-confusion (CoC) radii, reblurs the predicted sharp image, and uses the residual between the reblurred image and the input to correct the latent representation. This feedback branch is used only during training. Experiments show good restoration performance on DPDD and two cross-dataset evaluations, including strong results with single-step inference.

Critical strengths
  1. The motivation is clear and relevant to image restoration. By constraining the restored result through a forward imaging process, the authors attempt to address both visual sharpness and content fidelity. I find this direction meaningful.

  2. The CoC branch does not require additional depth or CoC annotations. Instead, it learns through reblurring supervision using sharp images, allowing the method to use existing blurry–sharp image pairs without extra annotations.

  3. The method achieves leading results under the reported evaluation settings. In addition to experiments on DPDD, the authors evaluate on LFDOF and RealDOF using a model trained only on DPDD, examining generalization across different devices and capture conditions. The reported distortion, perceptual quality, and observation consistency metrics also provide a relatively comprehensive evaluation.

Critical weaknesses
  1. The discussion of Dual-Pixel and Quad-Pixel defocus deblurring methods is insufficient. Although these methods use additional input information and do not necessarily need to be included in direct performance comparisons, they are closely related to the problem studied here. I suggest discussing their connections to and differences from the proposed method in the introduction or related work. Relevant examples include Learning dual-pixel alignment for defocus deblurring, Ldp: Language-driven dual-pixel image defocus deblurring network, and Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model.

  2. I am not fully convinced that reblurring error can adequately reflect the fidelity of recovered details. Defocus weakens or removes high-frequency information, so sharp images containing different details may produce very similar blurry observations. The lower reblurring error in Figure 5 indicates better consistency with the observation, but it does not establish that the recovered details match the original content. Text recognition, structural matching, or controlled high-frequency perturbation experiments would help support this claim.

  3. The test setting for A4 in Table 2 needs further clarification. The paper defines sharp discrepancy as the difference between the sharp ground truth and the endpoint prediction, but A4 also enables a correction based on this discrepancy during inference. It is unclear how this residual is obtained at test time. If sharp ground truth is used, this setting should be explicitly identified as an oracle comparison so that readers do not interpret it as a practical inference configuration.

  4. The model uses DINOv2-L, Qwen-Image VAE, and a relatively large LightningDiT. It is therefore difficult to determine how much of the improvement comes from the proposed method and how much comes from pretraining and model capacity. The ablations with the same backbone demonstrate the benefits of individual components, but substantial resource differences remain when comparing against smaller, task-specific deblurring networks. Strong baselines with comparable pretraining and computational budgets would make the comparison more convincing.

  5. The maximum CoC radius is set to 20 pixels, and the blur kernel radii follow a fixed, nonuniform sampling scheme, but the rationale for these choices is not sufficiently clear. Is a 20-pixel radius enough to cover severe defocus in the test data? How does the model handle blur beyond this range? I would like to see statistics on the blur range in the data and a sensitivity analysis of the radius limit and sampling scheme.

  6. Inference on images of arbitrary resolution needs to be described more clearly. The authors explain the token grids corresponding to 512-pixel crops and the 518-pixel DINO input, but do not specify whether testing uses tiling, how overlapping regions are merged, or how positional encodings are adapted to different image sizes. These details affect boundary quality, inference speed, and the fairness of comparisons with other methods.

  7. The evaluation settings in Table 1 are not sufficiently transparent. Do the baselines use official weights, or are they retrained or fine-tuned? Are the training data, iteration budgets, image resizing procedures, and metric implementations consistent? Given the relatively large PSNR improvements in some comparisons, I suggest providing reproducible evaluation configurations so that readers can assess whether the gains are obtained under comparable conditions. (Pages 7–8 and 13)

  8. The dataset order and numerical results in Section 4.2 are inconsistent. The text lists DPDD, LFDOF, and RealDOF, followed by 28.74, 29.88, and 31.29 dB. According to Table 1, however, the last two values correspond to RealDOF and LFDOF, respectively. The descriptions of the corresponding improvements should also be checked to avoid confusion about the cross-dataset results. (Pages 7–8)

Questions for authors

See Weakness

4 comments

Response to Reviewer vMQH (Part 1)

Thank you very much for taking the time to read our paper and for providing valuable feedback. We hope that our responses below adequately address your concerns and questions.

W1. Coverage of Dual-Pixel and Quad-Pixel methods

Response. Thank you for the suggestion. We have revised Sec. 2.1 (Single-Image Defocus Deblurring, p. 3, lines 129-136) to include discussions of DPANet [1], LDP [2], and QPDD/LMNet [3]. The revised discussion clarifies that these methods exploit dual-pixel or quad-pixel sub-aperture observations, whereas FCFlow operates on a conventional single RGB image. We further highlight the connection to LDP: both methods model spatially varying defocus, but LDP uses a blur map derived from dual-pixel observations to guide restoration, while FCFlow estimates an effective CoC map from a single RGB image and leverages differentiable reblurring feedback to enforce formation consistency during training. Accordingly, our quantitative comparisons follow the single-RGB-image input setting.

References

  1. Yu Li, Yaling Yi, Xinya Shu, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. Neurocomputing, 616:128880, 2025.
  2. Hao Yang, Liyuan Pan, Yan Yang, Richard Hartley, and Miaomiao Liu. LDP: Language-driven dual-pixel image defocus deblurring network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 24078–24087, 2024.
  3. Hang Chen, Yin Xie, Xiaoxiu Peng, Lihu Sun, Wenkai Su, Xiaodong Yang, and Chengming Liu. Quad-pixel image defocus deblurring: A new benchmark and model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5709–5719, 2025.

W2. Reblurring consistency and fidelity of recovered details

Response. Thank you for this important observation. We agree that lower reblurring error does not necessarily imply faithful detail recovery, as different sharp structures may produce similar blurred observations. Reblurring with the effective defocus forward model guides Flow Matching toward physically consistent restorations, while formation-consistent feedback helps align the recovered details with the original content. To address this concern, we have added a reference-based structural fidelity evaluation in Sec. 4.4, including quantitative boundary evaluation (Table 6) and visual comparisons (Fig. 6).

Specifically, using the existing checkpoints, we compared Vanilla FM (A0) and FCFlow (A7) at 1 NFE on all DPDD test pairs using Canny boundary matching with a 1-pixel tolerance. As shown below, FCFlow consistently improves boundary precision, recall, and mean Boundary F1 over Vanilla FM.

MethodPrecision ↑\uparrowRecall ↑\uparrowF1 ↑\uparrow
Vanilla FM (A0)0.79830.61030.6846
FCFlow (A7)0.86810.73050.7900

Fig. 6 further illustrates the improved boundary correspondence. These results provide complementary evidence of improved structural fidelity beyond reblurring consistency.

W3. Test-time sharp discrepancy in A4

Response. Thank you for pointing out this ambiguity. We clarify that A4 is an oracle ablation: its inference-time residual is computed using the paired sharp ground truth SS and the decoded endpoint prediction S~(t)\widetilde S^{(t)}, i.e., RA4=S−S~(t).R_{\mathrm{A4}}=S-\widetilde S^{(t)}. Since the sharp ground truth is unavailable in practical inference, A4 serves only as an oracle diagnostic rather than a deployable configuration. We have explicitly labeled A4 as an oracle correction in Table 2 (line 406) and clarified the residual definition in the “Test-time correction and inference NFE” paragraph of Sec. 4.3 (lines 482–483).

Response to Reviewer vMQH (Part 2)

W4. Contribution of model capacity and pretraining

Response. Thank you for raising this concern. To facilitate comparison across models with different capacities and pretraining, we report model complexity alongside restoration performance. All methods are profiled on the same NVIDIA RTX 4090 using 512×512512\times512 inputs for latency comparison. Results are summarized below.

MethodNFEParams (M)↓\downarrowFLOPs (T)↓\downarrowLatency (ms)↓\downarrowPSNR$\uparrow$LPIPS$\downarrow$
INIK-1.9770.648182.2426.14730.1815
NRK-6.0940.62931.2826.15190.2228
P2IKT-3.3190.68847.7226.33010.1920
ViTDeblur-121.6024.057621.3326.03640.1993
MBSD202529.53478.7324968.5326.56080.1066
FCFlow20839.72654.8421449.1029.34480.0920
FCFlow1839.7265.403156.8428.74320.1372

FCFlow uses 839.726M inference-loaded parameters, including DINOv2-L, Qwen-Image VAE, and LightningDiT. At 1 NFE, FCFlow achieves 28.74 dB PSNR with a latency of 156.84 ms. At 20 NFE, it achieves 29.34 dB PSNR. Compared with MBSD at the same NFE, FCFlow uses approximately one-third of the parameters, 69.7% of the FLOPs, and 29.2% of the inference time, while achieving a 2.78 dB higher PSNR.

To isolate the contribution of our training framework, we compare Vanilla FM and FCFlow using identical backbones and pretrained components. As reported in Table 2, FCFlow improves PSNR by 3.05 dB and 1.65 dB at 1 and 20 NFE, respectively, while reducing LPIPS, demonstrating the benefits of formation-aware training.

W5. CoC radius limit and kernel-bank sampling

Response. Thank you for raising these concerns. We conducted additional experiments on DPDD to investigate the radius sampling strategy and the 20-pixel upper bound. All configurations use 10 sampled radii. The results in Table 7 of the revised manuscript are reproduced below.

Variantrmax⁡r_{\max}SamplingPSNRLPIPS
A16 pxNon-uniform28.61270.1398
B (Ours)20 pxNon-uniform28.74320.1372
C30 pxNon-uniform28.73560.1375
D20 pxUniform28.65490.1390

With rmax⁡=20r_{\max}=20, non-uniform sampling outperforms uniform sampling, supporting denser sampling in the small-radius regime. Increasing rmax⁡r_{\max} from 16 to 20 pixels improves restoration, whereas extending it to 30 pixels brings no further benefit.

To characterize the effective blur-radius distribution, we analyzed 2,660 non-overlapping 256×256256\times256 DPDD blocks, selecting the disk radius that minimizes the RGB MAE between the reblurred sharp reference and observed blurred input. As shown in Fig. 8 of Appendix A.3, 91.20% of the blocks have best-fit radii within 20 pixels, while 8.80% exceed this range, including 4.25% reaching 30 pixels. Fig. 9 shows a representative region where the MAE is minimized at 20 pixels and increases for larger radii. These results support Cmax⁡=20C_{\max}=20 as a practical upper bound covering most, but not all regions. For regions exceeding this limit, the formation renderer provides only an approximation of the observed blur.

W6. Arbitrary-resolution inference

Response. Thank you for this suggestion. We have added detailed descriptions of the arbitrary-resolution inference procedure in Appendix A.4 (lines 885–902), including tiling, overlap fusion, and positional encoding.

Briefly, inference is performed at the original image resolution using 512×512512\times512 tiles with a 64-pixel overlap and 448-pixel stride. Boundary tiles are shifted to ensure full image coverage, and overlapping latent tiles are merged using normalized separable Hann weights before VAE decoding. Positional encodings are applied independently within each tile using fixed token grids (37×3737\times37 for DINOv2 and 32×3232\times32 for LightningDiT), without image-wide positional rescaling.

W7. Baseline weights and evaluation reproducibility

Response. Thank you for this comment. As clarified in Sec. 4.2 (lines 366–367), all baseline methods use their officially released pretrained checkpoints, without additional retraining or fine-tuning. Their training data and iteration budgets follow the respective official configurations. All methods are evaluated using the same test splits, image preprocessing settings, and metric implementations, as detailed in Appendix A.2.

W8. Dataset ordering and numerical inconsistency in Section 4.2

Response. We thank the reviewer for pointing out this inconsistency. The dataset ordering and corresponding numerical results have been corrected throughout the revised manuscript.

Thank you for the authors' response. The rebuttal has addressed most of my concerns, and I will revise my score accordingly.

We sincerely thank the reviewer for their follow-up and for acknowledging that our response addressed most of their concerns. We greatly appreciate their constructive feedback, which has helped us improve the paper.

Reviewer_4srzOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper puts a forward defocus model inside a latent flow-matching deblurring pipeline. A frozen DINOv2 encoder with a small adapter predicts a spatially varying CoC map directly from the blurred image, and that map drives a differentiable renderer built from ten disk PSFs with linear interpolation between neighbouring radii. There are no CoC labels; the map is learned by blurring the sharp target back to the observation. At training time the velocity at a random time is extrapolated one step to an endpoint, decoded, reblurred, and the residual against the observed blur produces a bounded latent correction, with the corrected endpoint supervised toward the sharp latent. At test time the branch is removed and sampling is plain FM. Training is on DPDD only, evaluation on DPDD, LFDOF and RealDOF, best at 20 steps, 28.74 dB at one step on DPDD.

Critical strengths

The results hold on three datasets whose acquisition differs a lot. Leading on LFDOF and RealDOF after training only on DPDD is not easy.

Using the branch in training and dropping it at test time is a more practical design than test-time data consistency. A6 vs A7 (28.7459 vs 28.7432) demonstrates that the formation branch can be dropped at test time without cost.

The A5 vs A7 comparison is well controlled and surprising. A5 feeds the module a residual that contains the ground truth; A7 does not. Everything else is identical, and A7 comes out 0.65 dB higher. This is the strongest piece of evidence for the paper's thesis.

Critical weaknesses
  1. LepL_{ep} reduces to LFML_{FM} when the corection is removed

With the correction set to zero, Eq. 13 reduces to Lep=‖z~s−zs‖1=t⋅‖vθ–(zb−zs)‖1L_{ep} = ‖z̃_s − z_s‖_1 = t·‖v_θ – (z_b-z_s)‖_1,

which is the Flow Matching objective of Eq. 6 with an L1 norm and a time weight. Two points follow that I would ask the authors to address. First, random-time endpoint optimization as described in contribution 2 is not a new supervisory signal when the feedback term is off. Second, with the feedback term enabled, Lep=‖t(vθ−(zb−zs))+λfbα⊙∆zfb‖1L_{ep} = ‖t(v_θ − (z_b-z_s)) + λ_{fb} α ⊙ ∆z_{fb}‖_1,

and since ∆max∆_{max}

and λfbλ_{fb} are both 1, the correction is allowed to be fairly large (compared with t(vθ−(zb−zs))t(v_θ − (z_b-z_s))), so part of the training loss could be carried by GφG_φ while vθv_θ itself is constrained only by LFML_{FM}. A6 against A7 differs by 0.0027 dB, which points in the same direction.

  1. The abstract over-credits the formation model The 1-NFE column of Table 2 decomposes as 0.21 (A1 vs A0) + 0.99 (A2 vs A1) + 1.20 (A5 vs A2) + 0.65 (A7 vs A5) = 3.05 dB, of which 0.86 dB (0.21 + 0.65) is specific to the formation model. Because A5 and A7 already isolate that part, I do not treat this as a flaw. But the abstract reads as though the whole gain belongs to the formation model, and a reader who adds up the table is likely to feel misled.
Questions for authors
  1. Could you freeze GφG_φ at its zero initialization so that the feedback term is off, while keeping LepL_{ep} and the schedule unchanged, and report the number? 200k steps would be enough to see the ordering. Besides, one more ablation with R(t) = 0 would separate the effect of model capacity from content.

  2. What is the typical value of ‖λfbα⊙∆zfb‖1/‖t(vθ−(zb−zs))‖1‖λ_{fb} α ⊙ ∆z_{fb}‖_1 / ‖t(v_θ − (z_b-z_s))‖_1 at the end of training?

  3. Does turning the adapter on at test time, with a larger ∆_max or λ_fb, help? If not, the correction framing is hard to sustain.

  4. A1 to A2 is +0.99 dB. What exactly is the A2 loss (image or latent domain, L1 weight, added on top of LFML_{FM}, or replacing it)? If the target is the same, where does the 1 dB come from?

  5. Training uses 512 crops with C_max = 20 px, but RealDOF is evaluated at 1680×1120. The CoC branch is gone at test time, yet the velocity field and DINOv2 never see that scale in training; the blur radius in pixels changes with resolution, while the 14×14 patch does not. Is any resampling or radius rescaling applied?

4 comments

Response to Reviewer 4srz (Part 1)

Thank you very much for taking the time to review our paper, and for providing many valuable comments and suggestions. We hope that the responses below adequately address your questions and concerns.

W1&Q1. Endpoint-loss equivalence and error absorption by correction

Response. Thank you for this insightful comment. We agree with the reviewer’s derivation. Along the linear interpolation path zt=(1−t)zs+tzbz_t=(1-t)z_s+t z_b, the uncorrected endpoint proposal satisfies

=t\big[(z_b-z_s)-v_\theta(z_t,t)\big].$$ Therefore, without feedback correction, the latent endpoint loss reduces to $$\mathcal L_{\mathrm{ep}}^{(t)} =t\|v_\theta(z_t,t)-(z_b-z_s)\|_1,$$ which is a time-weighted $L_1$ velocity-regression term. With feedback enabled, the endpoint loss is evaluated on the corrected proposal, allowing both the velocity field and feedback module to contribute to endpoint optimization. To determine whether the feedback also improves the learned velocity field, we trained a matched zero-correction control using the same training setup as A7, except that $\lambda_{\mathrm{fb}}=0$. Both models were evaluated on DPDD using flow-only inference without test-time correction. The results in **Table 3** of the revised manuscript are reproduced below. | Training variant | 1 NFE PSNR\$\uparrow\$ | 1 NFE LPIPS\$\downarrow\$ | 20 NFE PSNR\$\uparrow\$ | 20 NFE LPIPS\$\downarrow\$ | | --- | --- | --- | --- | --- | | Zero correction | 26.0147 | 0.1722 | 27.9503 | 0.1549 | | A7 (FCFlow) | **28.7432** | **0.1372** | **29.3448** | **0.0920** | | Improvement | +2.7285 | \$-\$0.0350 | +1.3945 | \$-\$0.0629 | Under flow-only inference, A7 consistently outperforms the zero-correction control, **improving PSNR by 2.73 dB at 1 NFE and 1.39 dB at 20 NFE**, while reducing LPIPS. These gains demonstrate that training-time feedback improves the learned velocity field rather than relying solely on explicit endpoint correction. ## W2. Attribution of the improvement to the formation model **Response.** Thank you for pointing this out. The 3.05 dB improvement reflects the cumulative gain of the complete FCFlow framework, rather than the formation model alone. As shown in **Table 2**, formation supervision (A1 vs. A0) and formation feedback over sharp feedback (A7 vs. A5) contribute 0.21 dB and 0.65 dB, respectively, totaling 0.86 dB. The remaining 2.19 dB comes from latent endpoint optimization and its integration with formation-consistent feedback. Hence, the proposed FCFlow framework incorporates an effective defocus forward model and formation feedback into the latent endpoint optimization within Flow Matching, improving restoration quality while preserving physical consistency. We have revised the abstract **(lines 16–17)** to clarify these contributions. ## Q1. Zero-feedback and zero-residual controls **Response.** Thank you for suggesting these additional controls. The zero-feedback condition is already evaluated in W1 by setting $\lambda_{\mathrm{fb}}=0$ while retaining latent endpoint supervision. Please refer to W1 for the quantitative results. To isolate the contribution of residual information, we additionally trained a zero-residual control by setting $R^{(t)}=0$ while keeping the feedback module, including its biases, trainable. Thus, the module can still produce nonzero corrections but receives no residual information. We used the same training configuration as A7. The results in **Table 3** of the revised manuscript are reproduced below. | Training variant | 1 NFE PSNR\$\uparrow\$ | 1 NFE LPIPS\$\downarrow\$ | 20 NFE PSNR\$\uparrow\$ | 20 NFE LPIPS\$\downarrow\$ | | --- | --- | --- | --- | --- | | Zero residual ($R^{(t)}=0$) | 25.9978 | 0.1687 | 27.8860 | 0.1588 | | A7 (formation residual) | **28.7432** | **0.1372** | **29.3448** | **0.0920** | | Improvement | +2.7454 | \$-\$0.0315 | +1.4588 | \$-\$0.0668 | **The zero-residual control performs comparably to the zero-correction control in W1, whereas A7 substantially outperforms both.** These results demonstrate that the performance gains rely on the formation information provided by the residual, rather than merely on the presence of a trainable feedback module.

Response to Reviewer 4srz (Part 2)

Q2. Correction-to-error ratio at the end of training

Response. Thank you for this question. We quantify the gated correction magnitude c=∣∣λfbα⊙Δzfb∣∣1c=||\lambda_{\mathrm{fb}}\alpha\odot\Delta z_{\mathrm{fb}}||_1, the uncorrected endpoint error e=∣∣z~s(t)−zs∣∣1e=||\widetilde z_s^{(t)}-z_s||_1, and their ratio r=c/(e+ϵ)r=c/(e+\epsilon), where ϵ=10−8\epsilon=10^{-8}. Using the final A7 checkpoint, we evaluated four interpolation times on paired DPDD test images. The interpolated states were constructed as zt=(1−t)zs+tzbz_t=(1-t)z_s+tz_b. All quantities are computed per sample and then averaged. We have added a detailed analysis of training-time feedback dynamics to Appendix A.5 of the revised manuscript. The results in Table 8 of the revised manuscript are reproduced below.

ttcceerr
0.1250.0035640.0153580.238072
0.3750.0035980.0468050.078607
0.6250.0036790.0806390.046481
0.8750.0037450.1211960.031496

As tt decreases, both the correction magnitude cc and endpoint error ee decrease. However, cc decreases only slightly, whereas ee decreases substantially more. Consequently, the correction-to-error ratio rr increases from 0.0315 at t=0.875t=0.875 to 0.238 at t=0.125t=0.125, indicating a larger relative contribution of feedback at earlier interpolation times.

Q3. Increasing correction strength at inference

Response. Thank you for raising this concern. Using the existing A6 checkpoint, we evaluated stronger test-time corrections by scaling the final gated latent update by 2×2\times, 4×4\times, and 8×8\times, without modifying the preceding Euler sampling trajectory or retraining the model. The results are summarized below.

Correction scale1 NFE20 NFE
1×1\times (A6)28.745929.3459
2×2\times28.737629.3234
4×4\times28.604229.1741
8×8\times28.393828.9147

Increasing the correction strength consistently degrades PSNR. Moreover, the default correction in A6 provides only negligible gains over flow-only inference in A7 (0.0027 dB at 1 NFE and 0.0011 dB at 20 NFE), whereas training-time formation feedback improves flow-only PSNR by 2.73 dB and 1.39 dB, respectively, as shown in W1. These results demonstrate that the primary benefit of formation feedback lies in shaping the learned velocity field during training, rather than explicit correction at inference, which can therefore be omitted.

Q4. Exact A2 objective and interpretation of its improvement

Response. Thank you for this important question. A2 introduces an additional image-domain (RGB-space) L1L_1 loss on the decoded, uncorrected endpoint proposal at a randomly sampled time t∼U(0,1)t\sim\mathcal{U}(0,1). It supplements, rather than replaces, the original FM loss and auxiliary reblurring loss. Specifically, the objective is

\mathcal L_{\mathrm{A2}} =&\mathcal L_{\mathrm{FM}} +\lambda_{\mathrm{rb}}\mathcal L_{\mathrm{rb}} +\lambda_{\mathrm{rgb}} \left\|D(\tilde z_s^{(t)})-S\right\|_1, \end{aligned}$$ where $\lambda_{\mathrm{rgb}}=1$, $D$ is the frozen VAE decoder, and $S$ is the paired sharp reference. Gradients from the RGB loss propagate through the decoder to the endpoint proposal and velocity field. Although both objectives use the same sharp reference, they supervise different quantities in different domains: the FM loss constrains the latent velocity prediction, whereas the RGB endpoint loss supervises the decoded restoration. Compared with A1, A2 improves PSNR by 0.9869 dB at 1 NFE and 0.7545 dB at 20 NFE on DPDD, demonstrating the benefit of supplementing latent-space flow matching with decoded-image endpoint supervision. ## Q5. Training crops, blur radius, and high-resolution evaluation **Response.** Thank you for raising this concern. We employ tiled inference for RealDOF evaluation. Specifically, the $1680\times1120$ images are processed using $512\times512$ tiles with a 64-pixel overlap and a 448-pixel stride. Boundary tiles are shifted as needed to ensure full image coverage, and overlapping predictions are fused using normalized Hann weights. Further details of the tiling and fusion procedure are provided in **Appendix A.4 (lines 885–902)**. Both training and inference operate on $512\times512$ tiles, preserving the same local processing field. Therefore, no additional tile resizing or CoC-radius rescaling is applied.

Thank the authors for the additional experiments and clarification comments. They addressed all my concerns in the last round of review.

We thank the reviewer for the prompt response and for taking the time to review our rebuttal. We are pleased that our response addressed your concerns. Thank you again for your constructive feedback and for helping us improve the paper.

Reviewer_s9F5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FCFlow, a training framework for single-image defocus deblurring based on latent blur-to-sharp Flow Matching. It uses the Qwen-Image VAE latent space and a LightningDiT velocity field conditioned on DINOv2 features. A CoC head predicts a spatially varying effective CoC map, supervised only by reblurring the sharp target through a differentiable disk-kernel renderer (Lrb\mathcal{L}{rb}). Trained on DPDD and evaluated on DPDD, LFDOF and RealDOF, FCFlow reports large gains (e.g., +2.18 dB PSNR over the best prior method on DPDD with 1 NFE), along with ablations on training components, NFE, PSF type and conditioning level.

Critical strengths

Strong, consistent gains at zero inference cost. FCFlow improves fidelity and perceptual metrics on all three datasets, including two held-out (cross-dataset) benchmarks. The same-backbone Vanilla FM control (A0) shows that much of the gain comes from the proposed training rather than from the backbone: about +3.05 dB at 1 NFE and +1.65 dB at 20 NFE on DPDD. Because the formation branch is dropped at test time, the improvement is free at inference, which is practically appealing.

Critical weaknesses

1.There is no evidence that the formation model shapes the final velocity field. The CoC head and the feedback branch are discarded at inference. Removing the correction changes PSNR by only 0.003 dB (A6 vs. A7), and Lrb\mathcal{L}{rb} alone adds just +0.01 dB at 20 NFE (A0→A1). The +1.8 dB gain over A2/A3 may therefore come from the latent-space endpoint loss itself, and controls for this (λfb=0\lambda{fb}=0, a wrong CoC) are missing. 2.The "decoupled from inference" claim is weak. Any training-only loss is decoupled from inference by construction. Also, the feedback branch requires no ground truth, yet it gives no benefit at test time, which suggests the formation discrepancy carries little usable information. This contradicts the paper's core motivation.

Questions for authors

1.What is the magnitude of the correction λfb,α⊙Δzfb\lambda_{fb},\alpha\odot\Delta z_{fb} relative to ∣z~s(t)−zs∣|\tilde z_s^{(t)} - z_s|, as a function of tt, during training and at test time? How does A7 compare with λfb=0\lambda_{fb}=0 under the same latent Lep\mathcal{L}_{ep}? This is the most important point for my assessment. 2.To show that the estimated CoC matters, could you report variants with (i) a constant or shuffled CoC map in the renderer, (ii) no renderer, feeding B−S~(t)B-\tilde S^{(t)} directly to GϕG_\phi, and (iii) stop-gradient on CC in the Lep\mathcal{L}{ep} path? Is CC currently detached in that path? If not, Lep\mathcal{L}{ep} also trains the CoC head, and CC may drift away from a physically meaningful CoC. 3.A4 claims inference-time correction with sharp feedback, whose input S−S~(t)S-\tilde S^{(t)} requires the ground truth. How is A4 computed at test time? For A4 and A6, is the correction applied at every Euler step or only to the final endpoint in 20-NFE inference? 4.Could you evaluate formation consistency with a formation model trained independently of FCFlow, or report restoration error stratified by CoC magnitude? This would support the claim that formation knowledge is transferred into the velocity field.

2 comments

Response to Reviewer s9F5 (Part 1)

Thank you very much for taking the time to review our paper. We also appreciate your insightful comments, questions, and valuable feedback. We hope that the responses below adequately address your concerns and questions.

W1&Q1&Q2. Evidence that formation feedback shapes the velocity field

Response. Thank you for this helpful comment. We agree that the original ablations did not fully isolate the contribution of formation feedback from that of latent endpoint supervision. To isolate the effect of training-time feedback, we trained an additional matched zero-correction control using the same training setup as A7, except that λfb=0\lambda_{\mathrm{fb}}=0. Both models were evaluated on DPDD using flow-only inference without test-time correction. The results in Table 3 of the revised manuscript are reproduced below.

Training variant1 NFE PSNR$\uparrow$1 NFE LPIPS$\downarrow$20 NFE PSNR$\uparrow$20 NFE LPIPS$\downarrow$
Zero correction26.01470.172227.95030.1549
A7 (FCFlow)28.74320.137229.34480.0920
Improvement+2.7285$-$0.0350+1.3945$-$0.0629

Compared with the zero-correction control, A7 improves PSNR by 2.73 dB at 1 NFE and 1.39 dB at 20 NFE, while reducing LPIPS by 0.0350 and 0.0629, respectively. These results show that training-time formation feedback provides additional benefits and contributes to shaping the learned velocity field.

We further investigate the role of the predicted CoC map in Q2 below. Shifting the predicted CoC map 5 pixels downward and rightward reduces PSNR by 2.20 dB at 1 NFE and 1.09 dB at 20 NFE compared with A7. This highlights the importance of the predicted CoC map in providing effective formation feedback.

W2. Significance of decoupling formation feedback from inference

Response. Thank you for this insightful comment. We agree that removing training-only branches at inference is not, by itself, a methodological contribution. Our central contribution is to exploit formation-consistent feedback during training to steer the velocity field toward physically consistent sharp-image restoration with fine details.

The zero-correction control (W1) demonstrates that training-time formation feedback improves the learned velocity field, while the test-time analysis (Q1) shows that explicit corrections remain small across trajectory states. These results support using formation feedback during training while omitting explicit correction at inference. We have revised the contribution statement in Sec. 1 (lines 115–117) accordingly.

Q1. Relative correction magnitude over time

Response. Thank you for this important question. We measure the gated correction magnitude c=∣λfbα⊙Δzfb∣1c=|\lambda_{\mathrm{fb}}\alpha\odot\Delta z_{\mathrm{fb}}|_1, the uncorrected endpoint error e=∣z~s(t)−zs∣1e=|\widetilde z_s^{(t)}-z_s|_1, and their ratio r=c/(e+ϵ)r=c/(e+\epsilon), where ϵ=10−8\epsilon=10^{-8}. All quantities are computed per sample and then averaged. We have added the corresponding analysis to Appendix A.5.

Training-time dynamics. We track intermediate checkpoints throughout the 1.4M-step training process at four interpolation times, t∈0.125,0.375,0.625,0.875t\in{0.125,0.375,0.625,0.875}, using paired DPDD images. As shown in Fig. 10, the correction magnitude initially increases and then stabilizes with small fluctuations, while the endpoint error progressively decreases. The statistics at the final checkpoint reported in Table 8, are reproduced below.

ttcceerr
0.1250.0035640.0153580.238072
0.3750.0035980.0468050.078607
0.6250.0036790.0806390.046481
0.8750.0037450.1211960.031496

The correction magnitudes remain similar across interpolation times, whereas the endpoint error increases with tt. The correction-to-error ratio decreases from 23.81% to 3.15%, indicating a larger relative contribution of feedback at earlier interpolation times.

Test-time dynamics. We evaluate potential corrections offline along the 20-NFE Euler trajectories of DPDD test pairs, without modifying the trajectories. The results reported in Table 9 are reproduced below.

ttcceerr
0.1250.0037300.1528840.024989
0.3750.0037300.1530750.024978
0.6250.0037320.1534450.024976
0.8750.0037360.1541050.024950
1.0000.0037370.1546120.024949

The correction magnitude remains 0.00373 and the endpoint error stays around 0.154, yielding a nearly constant correction-to-error ratio of approximately 2.49% across the evaluated states. This ratio is similarly small at t=1t=1, corresponding to the one-step endpoint proposal. These results indicate that endpoint correction contributes only marginally during inference, supporting one-step flow-only inference without the feedback branch.

Response to Reviewer s9F5 (Part 2)

Q2. CoC controls, feedback without rendering, and gradient routing

Response. Thank you for these important questions. In A7, the predicted CoC map is not detached, allowing gradients from the latent endpoint loss to propagate through the feedback renderer to the CoC head.

To investigate the contribution of the learned CoC map and its gradient pathway, we conducted three controlled ablations: (i) Shifted CoC: we shifted the predicted CoC map 5 pixels downward and rightward, retaining its values but changing their spatial locations; (ii) Direct residual: we replaced the rendered residual with B−S~(t)B-\tilde S^{(t)}, removing CoC-based reblurring from the feedback path; (iii) Stop-gradient CoC: we detached CoC only in the feedback renderer, preventing endpoint-loss gradients from propagating through the CoC-dependent rendering path. All variants use the same training configuration as A7. The results in Table 4 of the revised manuscript are reproduced below.

Feedback variant1 NFE PSNR1 NFE LPIPS20 NFE PSNR20 NFE LPIPS
Shifted CoC26.54030.171128.25940.1573
Direct residual26.25480.179328.05730.1621
Stop-gradient CoC26.64130.168928.34050.1604
A728.74320.137229.34480.0920

All three variants underperform A7 at both 1 and 20 NFE, with higher LPIPS. These results demonstrate that the learned CoC map provides useful information for formation feedback, and that its effectiveness benefits from CoC-based reblurring and gradient propagation through the feedback renderer.

Q3. Ground-truth access and correction schedule in A4 and A6

Response. Thank you for raising these questions. A4 is an oracle ablation that uses the paired sharp ground truth SS at test time to construct the feedback residual: RA4=S−D(zend)R_{\mathrm{A4}}=S-D(z_{\mathrm{end}}), where zendz_{\mathrm{end}} is the final latent endpoint before correction and DD is the VAE decoder. Thus, A4 requires ground-truth access and is not a deployable inference setting. We have labeled A4 as Sharp Feedback (w/ Oracle Correction) in Table 2 (line 406) and clarified its ground-truth-dependent residual in Sec. 4.3 (lines 482–483).

For 20-NFE inference, both A4 and A6 first complete all 20 Euler steps and then apply a single correction to the final latent endpoint. Specifically, A6 constructs its residual as $R_{\mathrm{A6}}=B-\mathcal{B}(D(z_{\mathrm{end}})) $ , where B\mathcal{B} denotes the CoC-based reblurring operator. The resulting gated correction is applied once to zendz_{\mathrm{end}} before final decoding.

Q4. Independent formation evaluation or performance by blur severity

Response. Thank you for this helpful suggestion. We conducted an additional evaluation stratified by CoC magnitude on the DPDD test set, using the existing Vanilla FM (A0) and A7 checkpoints without retraining. Specifically, we used the CoC maps predicted by the final A7 checkpoint and set two thresholds (4 and 12 pixels) based on their empirical distribution to divide pixels into mild, moderate, and severe blur regions. The resulting CoC maps and region masks were fixed and shared by both methods for a fair comparison.

For each image and CoC level, PSNR was computed over the corresponding region and averaged across images with valid pixels. The three regions contain 75, 76, and 76 valid images, respectively. All evaluations use flow-only inference without test-time correction. The results in Table 10 of the revised manuscript are reproduced below.

CoC levelRadius (px)Valid images1 NFE A01 NFE A71 NFE Δ\Delta20 NFE A020 NFE A720 NFE Δ\Delta
MildC<4C<47527.7028.59+0.8928.4829.10+0.62
Moderate4≤C<124\leq C<127628.4829.60+1.1229.4930.23+0.74
SevereC≥12C\geq127629.5531.13+1.5830.8431.92+1.08

FCFlow (A7) consistently outperforms Vanilla FM (A0) across all three CoC levels, where Δ\Delta denotes the PSNR gain of A7 over A0, with greater improvements in regions with stronger predicted defocus.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses to the Reviewers and Area Chairs

We sincerely thank all reviewers for their constructive feedback and the Area Chairs for coordinating the review process. We are encouraged that the reviewers recognize the well-motivated and physically grounded design of FCFlow for defocus deblurring (vMQH), its consistent restoration improvements and strong cross-dataset generalization across three benchmarks (s9F5, 4srz, vMQH), and the effectiveness of training-time formation feedback and the practical benefits of flow-only inference (s9F5, 4srz).

Our Contribution. The proposed FCFlow framework integrates an effective defocus forward model and formation-consistent feedback into latent endpoint optimization within Flow Matching. During training, formation discrepancies refine randomly sampled endpoint proposals, thereby shaping the learned velocity field toward more accurate and formation-consistent restoration. At inference, the formation-related branches are removed, retaining an efficient flow-only inference path without additional formation-related overhead.

Based on the reviewers' feedback, we have revised the manuscript and conducted additional experiments, with all changes highlighted in blue. The major revisions are summarized as follows:

  1. Related Work (Sec. 2.1). We added discussions of dual-pixel and quad-pixel defocus deblurring methods (DPANet, LDP, and QPDD/LMNet), clarifying their differences from our single-RGB-input setting (vMQH W1).

  2. Method Validation and Ablations (Sec. 4.3, Tables 2–4). We clarified the attribution of overall gains, identified A4 as an oracle-correction variant, and added matched zero-correction and zero-residual controls, along with CoC and feedback ablations. These results validate the effectiveness of training-time formation feedback in shaping the learned velocity field and the importance of CoC modeling (s9F5 W1/Q2, 4srz W1/Q1/W2, vMQH W3).

  3. Formation Consistency and Structural Fidelity (Sec. 4.4, Table 6 and Fig. 6). We added reference-based boundary evaluation to verify that improved reblurring consistency also corresponds to better structural fidelity (vMQH W2).

  4. Additional Analyses of Defocus Modeling and Feedback (Appendix A.3, Appendix A.5, Appendix B). We added CoC radius-range and sampling ablations, effective blur-radius statistics, analyses of training- and test-time feedback dynamics, and restoration evaluations stratified by predicted CoC severity. These analyses address concerns regarding the CoC upper bound, sampling design, feedback magnitude, and restoration performance under different defocus severities (s9F5 Q1/Q4, 4srz Q2, vMQH W5).

  5. Inference and Experimental Transparency (Sec. 4.2, Appendix A.2, Appendix A.4). We clarified the evaluation protocols and baseline checkpoint usage, and provided detailed arbitrary-resolution inference procedures with tiling and overlap fusion (4srz Q5, vMQH W4/W6/W7/W8).

We have made every effort to address the reviewers' concerns individually, with detailed point-by-point responses provided beneath each review for reference. We sincerely appreciate the reviewers' and the Area Chairs' time, thoughtful feedback, and valuable suggestions, which have helped us further strengthen this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.