acceptodds
Under review as a conference paper at ICLR 2027

Generative Refocusing: Flexible Defocus Control from a Single Image

Abstract

Depth-of-field control is essential in photography, but achieving perfect focus often requires multiple attempts or specialized equipment. Single-image refocusing is still difficult, requiring both sharp content recovery and realistic bokeh synthesis. Current methods have significant drawbacks: they require all-in-focus inputs, rely on synthetic data from simulators, and offer limited control over the aperture. We introduce Generative Refocusing, built around a training scheme that completes missing signals, converting synthetic renderings, single real bokeh photographs, and paired real captures into a unified training format. This allows us to scale training jointly across heterogeneous real-world sources that prior methods could not combine, letting our model learn authentic optical characteristics beyond what simulators capture while enabling precise control over focus and bokeh intensity. Our experiments show we achieve state-of-the-art performance in defocus deblurring, bokeh synthesis, and refocusing benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Model, Image Editing, Refocus, Defocus Deblur, Bokeh Synthesis

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_3Z7LOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GenRefocus, a two-stage refocusing framework. DeblurNet first recovers an all-in-focus image, and BokehNet then generates the result according to the target focal plane, defocus strength, and an optional aperture shape. A main contribution is completing missing supervision in different real-world data sources, allowing synthetic data, individual real images, and real paired data to be used jointly for training. The authors also construct a light-field-based refocusing benchmark.

Critical strengths
  1. The paper provides a concrete analysis of the training data problem. Different data sources lack different types of supervision. The authors combine deblurring, mask and depth estimation, EXIF information, and simulator calibration to convert these sources into a unified training format. This improves the usefulness of existing real-world data.

  2. The framework supports a relatively complete set of editing operations. Starting from an already defocused image, it can restore sharpness, adjust the focal plane, and control blur strength. Its scope is broader than methods that only adjust the aperture while keeping the focal plane fixed.

  3. The method performs well under the evaluation protocol used in the paper. With optimal blur strength search on LF-Bokeh, it achieves an LPIPS of 0.0816, outperforming the baselines listed in Table 3. It also obtains good reference-based perceptual results on LF-Refocus, demonstrating the effectiveness of the overall system.

  4. Aperture shape control is a meaningful extension. Using PointLight-1K and an additional LoRA, the model can adjust highlight bokeh according to a shape condition. I find this a useful control option for creative photography beyond simply adjusting blur strength.

Critical weaknesses
  1. The scene-level separation between training and test data needs to be described more clearly. The appendix explicitly states that the relevant EPFL samples in LF-Bokeh and LF-Refocus are excluded from training, but the deduplication between the remaining LIFF scenes and training sources such as LFDOF is less clearly explained. Page 24 also describes the entire LF-Bokeh dataset as unseen. I suggest making these descriptions consistent and providing scene-level splits and results on clearly held-out subsets so that readers can verify test set independence.

  2. The current evaluation mainly measures generation quality under optimal parameters and does not fully establish control accuracy in practical use. Both LF-Bokeh and LF-Refocus use the target ground truth to search for the blur strength that maximizes SSIM for each image. Applying the same search to baselines helps ensure fairness, but users cannot access the target image during actual editing. I would like to see experiments that select parameters without using the target ground truth and directly evaluate whether the outputs match the requested settings.

  3. The definition and calculation of LVCorr need clarification. The text describes it as the Pearson correlation between blur strength and Laplacian variance. However, stronger defocus usually decreases this variance, whereas Table 3 reports values close to positive 1. Is a sign reversal, absolute value, or another transformation applied? In addition, a high correlation only indicates a consistent trend; it does not directly measure the absolute control error of the target blur radius or focal plane.

  4. The construction of the DeblurNet training data is not fully described. The main text mentions approximately 3.5K paired samples, but Appendix C.1 only states that 3,000 of the sharpest images are selected from RealBokeh. It does not explain how the corresponding blurry inputs are selected, paired, or synthesized. Since this process directly determines the deblurring supervision, the complete data construction procedure should be provided. (Pages 6 and 15)

  5. The descriptions of the approximately 26K real training samples are not fully consistent between the main text and the appendix. The main text attributes them to ITW, RealBokeh, and LFDOF, while the appendix describes 13K existing ITW samples and 13K newly organized variable-aperture sequences. I suggest reporting the original count, paired-data proportion, and final count after deduplication or filtering for each source. It should also be clear which data sources receive manual mask refinement.

  6. The inference cost is high and may limit the practicality of interactive editing. According to the appendix, the default configuration takes approximately 143 seconds per image on an A6000. Even with only four steps, it still takes around 20.6 seconds. A runtime breakdown for the two generation stages, depth estimation, and tiled processing would be useful. Comparisons with fast bokeh and refocusing methods at the same resolution and on the same hardware would also clarify the trade-off between quality and speed.

Questions for authors

See weakness

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3Z7L (Part 1/3)

Thank you for recognizing the data-completion scheme and for identifying where the evaluation and training details need clarification.

W1 · Scene-level train/test separation

The scene-level separation between training and test data needs to be described more clearly. The appendix explicitly states that the relevant EPFL samples in LF-Bokeh and LF-Refocus are excluded from training, but the deduplication between the remaining LIFF scenes and training sources such as LFDOF is less clearly explained. Page 24 also describes the entire LF-Bokeh dataset as unseen. I suggest making these descriptions consistent and providing scene-level splits and results on clearly held-out subsets so that readers can verify test set independence.

The LIFF scenes in LF-Bokeh come exclusively from the official LFDOF test split, and all LFDOF samples used for training come from its official training split. The EPFL benchmark scenes are also held out from training. LF-Refocus inherits the same scene-level separation because its pairs are constructed from LF-Bokeh scenes. We have clarified these splits in the main text and appendix.

The table below reports results on the 30 held-out EPFL images in LF-Bokeh. GenRefocus achieves the lowest LPIPS (0.0902) and DISTS (0.0656), while BokehMe achieves the highest CLIP-I (0.9531 versus 0.9430 for GenRefocus).

EPFL-only results on LF-Bokeh

MethodLPIPS ↓DISTS ↓CLIP-I ↑
BokehMe0.10920.07550.9531
Bokehlicious0.17990.10630.9304
BokehDiff0.13930.09880.9444
Bokeh Diffusion0.51510.26680.6954
DiffCamera0.13110.08710.9237
MagicBokeh0.16480.12610.8873
GenRefocus0.09020.06560.9430

W2 · Evaluation of fidelity and controllability

The current evaluation mainly measures generation quality under optimal parameters and does not fully establish control accuracy in practical use. Both LF-Bokeh and LF-Refocus use the target ground truth to search for the blur strength that maximizes SSIM for each image. Applying the same search to baselines helps ensure fairness, but users cannot access the target image during actual editing. I would like to see experiments that select parameters without using the target ground truth and directly evaluate whether the outputs match the requested settings.

Our evaluation addresses two complementary aspects: perceptual fidelity under matched blur settings and the consistency of the blur response to K.

Following BokehMe (Peng et al., 2022), we search for each method’s optimal blur setting, using SSIM to the reference as the selection criterion. This accommodates differences in depth scales and blur controls across methods. To the best of our knowledge, existing real paired bokeh benchmarks do not jointly provide ground-truth K values and reliable metric depth for evaluating absolute control accuracy.

The LVCorr calculation itself does not require ground-truth images. Following DiffCamera (Wang et al., 2025a), we use LVCorr to complement the fidelity comparison by measuring whether increasing K at a fixed focal plane consistently increases the generated blur, which is a behavior directly relevant to practical editing. It provides evidence of control consistency beyond image fidelity alone, but does not measure absolute blur-radius accuracy.

W3 · LVCorr calculation

The definition and calculation of LVCorr need clarification. The text describes it as the Pearson correlation between blur strength and Laplacian variance. However, stronger defocus usually decreases this variance, whereas Table 3 reports values close to positive 1. Is a sign reversal, absolute value, or another transformation applied? In addition, a high correlation only indicates a consistent trend; it does not directly measure the absolute control error of the target blur radius or focal plane.

We compute LVCorr as the Pearson correlation between K and the negative Laplacian variance of grayscale outputs within each fixed scene and focal plane, then average over valid groups following previous research. Stronger blur typically reduces Laplacian variance, so this sign reversal makes the expected correlation positive. We have clarified this in the revised manuscript.

We agree that LVCorr measures the linear consistency of the blur response rather than absolute errors in blur radius or focal-plane placement. It complements the perceptual metrics, allowing us to assess both perceptual fidelity and control consistency.

AuthorsOpenReview ↗

Response to Reviewer 3Z7L (Part 2/3)

W4 · DeblurNet training pairs

The construction of the DeblurNet training data is not fully described. The main text mentions approximately 3.5K paired samples, but Appendix C.1 only states that 3,000 of the sharpest images are selected from RealBokeh. It does not explain how the corresponding blurry inputs are selected, paired, or synthesized. Since this process directly determines the deblurring supervision, the complete data construction procedure should be provided. (Pages 6 and 15)

Thank you for pointing out. The manuscript incorrectly reported the number of training pairs. We select approximately 3,000 sharp references from RealBokeh, yielding approximately 15K blurred/sharp pairs rather than 3,000 pairs.

Each RealBokeh aperture series captures the same scene at a fixed focal plane. We rank the smallest-aperture references (f/22) by Laplacian variance and retain the top 75% (approximately 3,000 references). Each reference is paired with all available images at other aperture settings from the same series as blurry inputs. Together with the native pairs from the DPDD training split, this gives approximately 15.5K training pairs. All pairs consist of real captures; no blur synthesis is involved. We haved corrected the sample count and clarified the complete pairing procedure.

W5 · Real BokehNet training data and mask annotation

The descriptions of the approximately 26K real training samples are not fully consistent between the main text and the appendix. The main text attributes them to ITW, RealBokeh, and LFDOF, while the appendix describes 13K existing ITW samples and 13K newly organized variable-aperture sequences. I suggest reporting the original count, paired-data proportion, and final count after deduplication or filtering for each source. It should also be clear which data sources receive manual mask refinement.

We have aligned the descriptions in the main text and the supplementary. “Newly curated” refers to our reorganization and annotation of existing datasets, rather than newly captured images.

ITW. We use approximately 13K non-all-in-focus images selected by the original authors and directly use the BiRefNet focus masks.

RealBokeh. We inspect 3.7K candidate reference images from the training split and exclude 1K whose focus regions remain blurry, retaining 2.7K references. Around 30% of the retained references use manually annotated focus masks. We select images at 3–5 aperture settings per scene and exclude excessively blurred targets, yielding approximately 12K pairs.

LFDOF. We use a subset of 113 scenes from the LFDOF training split and manually annotate 300 focus masks for the selected reference images. Each reference is paired with 2–4 bokeh images, yielding 1K pairs.

Together, RealBokeh and LFDOF contribute approximately 13K pairs, accounting for about half of the approximately 26K real training examples. We have reported these details consistently in revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 3Z7L (Part 3/3)

W6 · inference cost

The inference cost is high and may limit the practicality of interactive editing. According to the appendix, the default configuration takes approximately 143 seconds per image on an A6000. Even with only four steps, it still takes around 20.6 seconds. A runtime breakdown for the two generation stages, depth estimation, and tiled processing would be useful. Comparisons with fast bokeh and refocusing methods at the same resolution and on the same hardware would also clarify the trade-off between quality and speed.

We report both initial processing time and subsequent editing latency, together with comparisons at the same resolution and on the same GPU.

For a 512×512 image on a single NVIDIA RTX A6000, non-tiled inference with 28 denoising steps per generation stage takes approximately 18 s for DeblurNet, around 1 s for depth estimation, and 27 s for BokehNet, totaling 46 s. The all-in-focus image and depth map can be cached, so subsequent changes to the focal point or K require only BokehNet.

Stage-wise runtime on LF-Refocus images at 1008×688 resolution. All measurements use a single NVIDIA RTX A6000; the step count applies to each generation stage. BokehNet time also represents subsequent editing latency when the all-in-focus image and depth map are cached.

StepsTilingDeblurNet (s)Depth Pro (s)BokehNet / edit (s)Initial total (s)
28Yes108.5381.056143.371252.965
4Yes15.8541.06420.55637.474
28No42.5481.04863.263106.859
4No6.3451.0539.40516.803

With four steps and no tiling, initial processing takes approximately 16.8 s, while each subsequent edit takes approximately 9.4 s. These latency measurements are separate from the 28-step tiled quality results reported below.

Tiled inference incurs additional computation by processing multiple overlapping tiles, with runtime increasing with the number of tiles. In our additional in-the-wild tests, images up to 1500×1500 pixels could be processed without tiling.

Comparison with faster methods. All methods below are evaluated at 1008×688 on a single NVIDIA RTX A6000. Our results use 28-step tiled inference.

Bokeh synthesis:

MethodLPIPS ↓DISTS ↓CLIP-I ↑Time (s)
BokehMe0.12600.07460.95050.052
Bokehlicious0.21600.10330.93640.124
BokehDiff0.17730.09120.94751.419
Bokeh Diffusion0.48900.24560.758388.732
MagicBokeh0.19280.10740.92121.573
GenRefocus0.08160.04900.9711143.371

End-to-end refocusing:

MethodLPIPS ↓DISTS ↓CLIP-I ↑Time (s)
Bokehlicious deblurring → Depth Pro → BokehMe0.15020.09150.93771.584
MagicBokeh0.20370.11550.90491.505
GenRefocus0.13240.07590.9571252.965

Our 28-step tiled configuration achieves the best LPIPS, DISTS, and CLIP-I among the compared methods in both tasks. For lower editing latency, users can reduce the denoising steps and disable tiling, while caching avoids repeating deblurring and depth estimation.

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

This paper presents a Generative Refocussing approach that take a single image and allows it to be refocussed in different ways. Their system consists of two modules. A DeblurNet which is a LoRA adapted diffusion-based FLUX system and a BokehNet which is also a LoRA adapted FLUX system (from what I can figure out). The method can use different types of inputs available: synthetic images, single images with bokeh, paired images of AIF-blurred pairs, etc. The DeblurNet converts any image into an all-in-focus image and recovers other parameters. The processing is done in an VAE emcoded space. BokehNet needs as input the AIF image, defocus map and optionally aperture shape. These are encoded and concatenated and presented to the BokehNet for iterative denoising.

The paper discusses different datasets, their limitations, and ways to adapt them for training their refocussing system. The main paper discusses some and the appendix mentions many more.

Quantitative results are shown on a few datasets and against several prior methods and presented in several tables in the paper as well as the appendix. The proposed method performs better in most cases. Qualitative results are shown on several images from standard datasets as well as on a few historical images and in-the-wild images. User-study results are also shown for some of the results against a couple of prior methods.

Critical strengths

An elaborate scheme to create appropriate data for training such a system is presented. While many of the methods are based on heuristics like finding the closest-images using binary search, etc., the authors seem to have created a substantial core of images to train their system. Some of those are created with substantial manual involvement also. I assume the images and combinations created using such a process will be made available to the community for furthering research in related areas. I couldn't see it clearly mentioned anywhere, however.

I mention the training data specially as the architecture of the DeblurNet and BokehNet otherwise are quite standard, being built as adaptations of the FLUX system. The results seem good qualitatively.

Critical weaknesses

I have a few observations and comments about the work and its presentation in the manuscript.

  • I found the description to be difficult to follow as several tricks have been used to create the dataset for training based on what source was used. A clear compilation of the different ways of transforming available images into useful training data will greatly enhance the understanding of the readers.
  • I am still not clear how a new image (say, a wild one) is refocussed using this system. A step-by-step description of that will help understand the method. What is the typical inference time or time to generate a newly focussed image? Appendix O talks about 143 seconds, but is that typical? Mentioning the training and generation time in the main paper will help greatly.
  • I find the discussion on the aperture shape somewhat orthogonal. I don't see it as a significant contribution; it can be an additional feature. Nothing wrong with it, but I don't see very clear validation of it in the results either.
  • While it is nice to see many examples and comparisons in the supplementary, I must observe almost all images given there are relatively simple, with most having a foreground object and a background. There were not many examples of continuous change of focus there. All the refocussing examples in the wild and historical categories are simple in this manner. Given the title of the paper includes "DoF Rendering", I don't see sufficient demonstration of that on arbitrary images.
  • In my view, the curation of diverse datasets available into a form usable for refocussing research is a significant contribution of this work, provided the results are usable for future research. This part is not clearly mentioned in the paper anywhere.
Questions for authors

Please see the weakness section and address those concerns!

2 comments
AuthorsOpenReview ↗

Response to Reviewer hzDU (Part 1/2)

We thank the reviewer for recognizing the value of our training data curation and the quality of our qualitative results.

W1 · Training data preparation

I found the description to be difficult to follow as several tricks have been used to create the dataset for training based on what source was used. A clear compilation of the different ways of transforming available images into useful training data will greatly enhance the understanding of the readers.

Thank you for the suggestion. We have added a source-by-source summary to Sec. 3.2. Each source supplies the same training triplet: an all-in-focus input, a target bokeh image, and a defocus map.

Synthetic data. We start from sharp images, estimate depth, sample the focal plane and K, and render the target with BokehMe.

Single real bokeh photos. We keep the captured bokeh image as the target, estimate its all-in-focus counterpart with DeblurNet, infer the focal plane from the focus mask and inverse depth, and derive K from EXIF.

Paired real captures. We retain the image pairs, estimate the focal plane from a verified or refined focus mask, and fit K by matching simulated bokeh to the captured target. We filter unreliable examples.

In all three cases, we compute the defocus map from the focal plane, depth map, and K to complete the same training triplet.

W2 · Inference workflow and runtime

I am still not clear how a new image (say, a wild one) is refocussed using this system. A step-by-step description of that will help understand the method. What is the typical inference time or time to generate a newly focussed image? Appendix O talks about 143 seconds, but is that typical? Mentioning the training and generation time in the main paper will help greatly.

Consider an in-the-wild image in which the user clicks on a position they want to bring into focus and selects a blur strength K. We refocus the image as follows.

  1. Recover an all-in-focus image. DeblurNet estimates an all-in-focus version of the input image.

  2. Estimate scene depth. Depth Pro predicts a depth map from the recovered image.

  3. Construct the defocus map. We take the inverse depth at the clicked point as the target focal plane and compute the defocus map using the selected K.

  4. Render the refocused image. BokehNet takes the all-in-focus image and the defocus map as conditions and generates the result.

Inference time. refocusing a 512×512 image on a single NVIDIA RTX A6000 without tiling takes 46 s with 28 denoising steps per generation stage: 18 s for DeblurNet, approximately 1 s for depth estimation, and 27 s for BokehNet.

The 143.4 s per image reported in Appendix O was measured on BokehNet using tiled inference with 28 denoising steps. Under the same tiled protocol, reducing the step count to 10 or 4 brings the runtime to 50.8 s or 20.6 s per image, respectively, with only modest changes in perceptual fidelity (Table 11). Without tiling, BokehNet takes approximately 63 s at 28 steps at 1008×688 resolution, providing a 2.3× speedup. Non-tiled inference or fewer denoising steps offers a lower-latency option for everyday editing.

Training time. on two NVIDIA RTX A6000 GPUs, DeblurNet took approximately 3 days to train for 60K steps, and BokehNet took approximately 5 days to train for 100K steps.

We have reported the training and inference times in the main text.

W3 · Aperture-shape control

I find the discussion on the aperture shape somewhat orthogonal. I don't see it as a significant contribution; it can be an additional feature. Nothing wrong with it, but I don't see very clear validation of it in the results either.

We agree that aperture-shape control is an additional feature. The current validation consists of qualitative examples of triangle, heart, and star bokeh following synthetic fine-tuning. We will present it as an extension and identify the development of real paired datasets with varied aperture shapes as a direction for future training and evaluation.

W4 · Continuous focus sweeps

While it is nice to see many examples and comparisons in the supplementary, I must observe almost all images given there are relatively simple, with most having a foreground object and a background. There were not many examples of continuous change of focus there. All the refocussing examples in the wild and historical categories are simple in this manner. Given the title of the paper includes "DoF Rendering", I don't see sufficient demonstration of that on arbitrary images.

Thank you for raising this point. The first row of the HTML supplement contains several videos demonstrating continuous change of focus. In particular, videos 2, 7, 8, and 10 (counting from left to right) show scenes with multiple people at different distances from the camera. Each demo starts from a photograph with one person in focus and shows our method smoothly shifting focus between people at different depths.

AuthorsOpenReview ↗

Response to Reviewer hzDU (Part 2/2)

W5 · Release plan for the curated data

In my view, the curation of diverse datasets available into a form usable for refocussing research is a significant contribution of this work, provided the results are usable for future research. This part is not clearly mentioned in the paper anywhere.

We thank the reviewer for recognizing the value of our dataset curation. We plan to release the curated data, benchmarks, and training code upon acceptance for future research.

Reviewer_EiP8OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes GenRefocus, a single-image refocusing framework with two stages built on a shared FLUX-based architecture:

  • DeblurNet recovers an all-in-focus image from an input in any focus state.
  • BokehNet renders bokeh from a defocus map D_def = K·|D − D_focus|, with an optional aperture-shape condition.

A "data-completion" scheme pseudo-labels the missing AIF image, focus plane, and K, so that synthetic renderings and real bokeh data can be trained jointly. Two light-field benchmarks (LF-Bokeh, LF-Refocus) are introduced.

Critical strengths
  1. Unified framework. Defocus deblurring and bokeh rendering share one generative architecture and conditioning design. This enables refocusing from inputs in any focus state, at arbitrary resolution, with control over focus plane, blur strength, and aperture shape. App. N shows the two-stage design beats direct refocusing.
  2. Promising results. The method achieves the best results on DPDD deblurring, LF-Bokeh fidelity (LPIPS 0.082 vs. 0.124 for BokehMe), and LF-Refocus. It is also strongly preferred in the user studies, and its in-the-wild and historical-photo results are convincing.
Critical weaknesses
  1. K is not intuitive and lacks clear physical meaning.
    • K folds focal length, f-number, focus distance, and resolution into one gain on inverse depth.
    • This is acceptable for editing, where K works as a relative blur slider. However, the units should be clarified: App. M (px·mm) implies values around 10⁴–10⁵, while the figures use 5–150.
    • The paper should also explain how EXIF-derived K (ITW) and simulator-fitted K (paired data) are put on a common scale, since K is never directly supervised.
  2. Evaluation.
    • Fidelity relies on a per-image oracle search over K, and LVCorr measures only monotonicity.
    • LF-Bokeh GT comes from small-baseline Lytro light fields, which lack real-lens effects. 170 of its 200 images are from LIFF, but only the EPFL subset is stated to be held out. Please confirm there is no overlap with training data and report EPFL-only results.
    • PSNR/SSIM for deblurring are missing.
  3. Data description. Sec. 3.2/4.1 attribute the paired real data to LFDOF and RealBokeh, but App. C.2 describes 13K "newly curated" aperture series. The K sampling range and the SSIM threshold are also not reported.
Questions for authors
  1. What normalization of D and D_def does BokehNet receive? How do the figures' K values relate to App. M?
  2. How are synthetic-data edge and occlusion artifacts handled? Is the full BokehMe (classical + neural) used?
  3. Is there overlap between LIFF/Lytro training data and the LF benchmarks?
  4. What is the source of the 13K curated series? Will they and the benchmarks be released?
2 comments
AuthorsOpenReview ↗

Response to Reviewer EiP8 (Part 1/2)

We thank the reviewer for the constructive feedback and for recognizing the strengths of our unified two-stage framework and the results on deblurring and refocusing.

W1-1 / W1-2 / Q1 · Physical meaning, units, and normalization

K is not intuitive and lacks clear physical meaning.K folds focal length, f-number, focus distance, and resolution into one gain on inverse depth.This is acceptable for editing, where K works as a relative blur slider. However, the units should be clarified: App. M (px·mm) implies values around 10⁴–10⁵, while the figures use 5–150.

What normalization of D and D_def does BokehNet receive? How do the figures' K values relate to App. M?

Physical meaning of K. K specifies how many pixels the circle-of-confusion radius increases for each 1m−11\mathrm{m}^{-1} increase in the absolute inverse-depth difference from the focal plane.

Units in Appendix M and the figures. Appendix M contains a unit error. We express depth in meters, so the correct unit of K is px·m. With this correction, the K in Appendix M is on the same scale as the values shown in the figures. We have corrected the units in Appendix M.

Normalization of D and D_def. In implementation, we take the reciprocal of the depth map in meters to obtain D = 1/z, without any additional normalization. Before VAE encoding, we divide D_def by a fixed constant (100), clip the result to [0,1], and replicate it across three channels.

W1-3 · Common scale for EXIF-derived and fitted K

The paper should also explain how EXIF-derived K (ITW) and simulator-fitted K (paired data) are put on a common scale, since K is never directly supervised.

For both data sources, we resize images to a long edge of 1024 pixels while preserving their aspect ratio, and use metric depth estimates from Depth Pro in meters. For ITW, we compute the pixel ratio at the resized resolution and use it to derive K from EXIF metadata. For paired data, we fit K with the simulator at the same resolution. Thus, for a given inverse-depth difference, the same K corresponds to the same circle-of-confusion radius in pixels for both sources.

W2-1 · Fidelity and controllability

Fidelity relies on a per-image oracle search over K, and LVCorr measures only monotonicity.

The same K setting can produce different blur radii across methods. We therefore search over each method’s own blur control for each image and select the output with the highest SSIM to the reference following BokehMe (Peng et al., 2022). This gives each method the same opportunity to match the reference blur in the fidelity comparison. LVCorr complements the perceptual metrics by measuring how consistently the generated blur increases with K at a fixed focal plane.

W2-2 / Q3 · Benchmark independence and real-lens effects

LF-Bokeh GT comes from small-baseline Lytro light fields, which lack real-lens effects. 170 of its 200 images are from LIFF, but only the EPFL subset is stated to be held out. Please confirm there is no overlap with training data and report EPFL-only results.

Is there overlap between LIFF/Lytro training data and the LF benchmarks?

Scene-level independence. There is no scene-level overlap between our training data and the benchmarks. The LIFF portion of LF-Bokeh comes exclusively from the LFDOF test split, and all LFDOF samples used for training come from its official training split. EPFL benchmark scenes are also held out. LF-Refocus inherits the same scene split because its pairs are built from LF-Bokeh scenes. We have clarified this in the main text and appendix.

The table below reports results on the 30 held-out EPFL images in LF-Bokeh. GenRefocus achieves the lowest LPIPS (0.0902) and DISTS (0.0656) on this subset.

EPFL-only results on LF-Bokeh

MethodLPIPS ↓DISTS ↓CLIP-I ↑
BokehMe0.10920.07550.9531
Bokehlicious0.17990.10630.9304
BokehDiff0.13930.09880.9444
Bokeh Diffusion0.51510.26680.6954
DiffCamera0.13110.08710.9237
MagicBokeh0.16480.12610.8873
GenRefocus0.09020.06560.9430

Real-lens effects. We acknowledge that LF-Bokeh has limited coverage of real-lens effects. For further evaluation of bokeh realism, please refer to our user study (Table 6) and RealBokeh comparison (Table 12).

AuthorsOpenReview ↗

Response to Reviewer EiP8 (Part 2/2)

W2-3 · PSNR and SSIM for deblurring

PSNR/SSIM for deblurring are missing.

We report PSNR and SSIM on RealDOF and DPDD below.

PSNR and SSIM for defocus deblurring

MethodRealDOF PSNR ↑RealDOF SSIM ↑DPDD PSNR ↑DPDD SSIM ↑
DRBNet25.74000.771025.74500.7910
Restormer25.08140.767925.97680.8102
INIKNet25.30000.761026.05800.8010
Bokehlicious25.80000.790525.91150.8058
DiffCamera21.94230.600323.24820.6634
GenRefocus23.43400.706524.27740.7316

GenRefocus achieves lower PSNR and SSIM than the strongest baselines but leads on several perceptual metrics in Table 2, a pattern consistent with the perception–distortion trade-off (Blau & Michaeli, 2018). We will revise the abstract and introduction to describe our gains in terms of perceptual quality.

W3 / Q4 · Data provenance, release, and filtering

Data description. Sec. 3.2/4.1 attribute the paired real data to LFDOF and RealBokeh, but App. C.2 describes 13K "newly curated" aperture series.

The K sampling range and the SSIM threshold are also not reported.

What is the source of the 13K curated series? Will they and the benchmarks be released?

Data provenance and release. The approximately 13K paired examples were curated from existing RealBokeh and LFDOF captures. We screened the examples, verified or refined focus masks, and estimated the blur-strength labels. We will name both sources explicitly in Appendix C.2. We plan to release the curated data and benchmarks upon acceptance.

K range and SSIM filtering. For the paired real data, the K sampling range is 0–200 px·m, and we retain the pair only if its best-fit SSIM exceeds 0.85. This threshold removes approximately 10% of the candidate pairs. We have reported the search range and threshold in the revised manuscript.

Q2 · Synthetic edge and occlusion artifacts

How are synthetic-data edge and occlusion artifacts handled? Is the full BokehMe (classical + neural) used?

We use the full BokehMe pipeline, including its neural refinement stage. The refinement mitigates edge and occlusion artifacts, though some artifacts remain. We use synthetic data to pretrain BokehNet on focus and blur control, then fine-tune on real bokeh images to reduce reliance on simulator-specific artifacts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.