- 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
-
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.
-
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.
-
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.
-
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
-
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 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.
-
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.
-
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)
-
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.
-
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
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.
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.
| Steps | Tiling | DeblurNet (s) | Depth Pro (s) | BokehNet / edit (s) | Initial total (s) |
|---|---|---|---|---|---|
| 28 | Yes | 108.538 | 1.056 | 143.371 | 252.965 |
| 4 | Yes | 15.854 | 1.064 | 20.556 | 37.474 |
| 28 | No | 42.548 | 1.048 | 63.263 | 106.859 |
| 4 | No | 6.345 | 1.053 | 9.405 | 16.803 |
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:
| Method | LPIPS ↓ | DISTS ↓ | CLIP-I ↑ | Time (s) |
|---|---|---|---|---|
| BokehMe | 0.1260 | 0.0746 | 0.9505 | 0.052 |
| Bokehlicious | 0.2160 | 0.1033 | 0.9364 | 0.124 |
| BokehDiff | 0.1773 | 0.0912 | 0.9475 | 1.419 |
| Bokeh Diffusion | 0.4890 | 0.2456 | 0.7583 | 88.732 |
| MagicBokeh | 0.1928 | 0.1074 | 0.9212 | 1.573 |
| GenRefocus | 0.0816 | 0.0490 | 0.9711 | 143.371 |
End-to-end refocusing:
| Method | LPIPS ↓ | DISTS ↓ | CLIP-I ↑ | Time (s) |
|---|---|---|---|---|
| Bokehlicious deblurring → Depth Pro → BokehMe | 0.1502 | 0.0915 | 0.9377 | 1.584 |
| MagicBokeh | 0.2037 | 0.1155 | 0.9049 | 1.505 |
| GenRefocus | 0.1324 | 0.0759 | 0.9571 | 252.965 |
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.
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 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
W2 · Evaluation of fidelity and controllability
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
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.