- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
LensMaster formulates depth-of-field editing as a transformation between source and target optical states. It unifies multiple editing tasks through focal and aperture conditions, identity, cycle, and composition consistency constraints, and spatially varying lens feature modulation. The authors also collect approximately 40,000 FujiBokeh images and train the model jointly with RealBokeh. The method is evaluated on bokeh synthesis, defocus removal, and refocusing.
Critical strengths
-
Conditioning on both source and target states is a reasonable design. Inputs with different degrees of defocus may require different restoration processes even when the target aperture and focal setting are the same. Using the focal and aperture settings at both ends helps describe aperture opening and closing, refocusing, and joint editing within a unified framework. (Pages 3–4, Equations 1–6)
-
The method does not require separately estimated depth, segmentation, or CoC maps. This reduces its dependence on intermediate estimates and provides an alternative approach to end-to-end depth-of-field editing. (Pages 3–4)
-
The experiments cover a broad range of editing tasks. Tables 2–4 evaluate bokeh synthesis from all-in-focus images, defocus removal, and four types of arbitrary-state editing. Table 5 further examines source conditioning, consistency constraints, and lens modulation. The results show relatively consistent performance gains across these settings.
Critical weaknesses
-
Missing implementation and training details are a major obstacle to reproducibility. The main text only broadly mentions a pretrained autoencoder, a Transformer, and conditional flow matching. It does not specify the backbone, pretrained weights, model size, number of training epochs, optimizer, sampling steps, or loss weights. Of the 12 pages provided, the last three are references, and a complete implementation configuration is still missing. These details should be included.
-
The implementation of the consistency losses within the flow matching framework needs a more concrete explanation. Equations 14–16 treat the generator as a directly callable mapping, but generation typically involves multi-step integration, and the cycle and composition constraints require repeated generator calls. Do the authors backpropagate through the full sampling process, use a single-step endpoint approximation, stop gradients, or introduce a teacher branch? These choices substantially affect both training cost and the meaning of the constraints, and I cannot determine them from the current equations.
-
It is unclear how random noise is handled across different generation paths. The model starts from Gaussian noise, but the identity, cycle, and composition consistency equations do not explicitly include random variables. If different paths use independent noise, pixelwise differences may reflect both editing errors and stochastic generation differences. If noise is shared, the sharing scheme should be specified. I suggest describing the noise settings and evaluating sensitivity to random seeds.
-
The current experiments do not sufficiently establish the model's lens modeling capability. The quantitative results mainly report task performance separately on RealBokeh and FujiBokeh. There is no one-to-one comparison with real captures of the same scene under different target lenses, nor an evaluation on unseen lenses. The evidence therefore more clearly supports adaptation to two seen data domains, while cross-lens transfer still requires additional validation.
-
The final configurations and evaluation tasks in the different ablation groups of Table 5 are unclear. The last rows of the three groups report 27.68/0.1548, 27.68/0.1648, and 28.19/0.1326, respectively, without explaining whether they correspond to different tasks, data subsets, or fixed components. If all three rows represent the full model, the differences should be explained. If the configurations differ, they should be listed directly in the table to make the comparisons easier to follow. (Page 9, Table 5)
-
The individual contributions of the consistency constraints are not sufficiently separated. Table 5 adds the identity and cycle losses together, making it difficult to determine how much benefit comes from the cycle constraint itself. I suggest evaluating the identity, cycle, and composition losses separately and including a sensitivity analysis of their weights. Identity error, round-trip editing error, and drift after repeated edits would also provide more direct evidence that these constraints achieve their intended goals. (Pages 5 and 9)
-
The claim of recovering extremely defocused regions needs more evidence of content fidelity. The examples show that the model can generate visually plausible details, but whether those details match the original structures remains unclear. I would like to see targeted evaluations on text and thin lines, along with failure cases involving transparent objects, reflections, occlusions, and inputs from unknown lenses. These experiments would help clarify the method's limitations.
-
The reporting of computational cost and result stability is incomplete. The paper does not provide inference latency, peak memory usage, or training cost. It also lacks variation across random seeds or confidence intervals computed over test scenes. Given the multiple generation calls and regularization terms in this framework, these details are important for judging whether the performance gains justify the additional cost and whether small differences in the reported metrics are stable. (Pages 5–9)
Questions for authors
See weakness