acceptodds
Desk rejected at ICLR 2027

TextBokeh: Language-Guided Depth-of-Field Editing with Target and Strength Control

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Depth of field controls which subject appears sharp and how strongly the rest of a scene is defocused. Existing post-capture interfaces commonly require a click, mask, focal distance, or numerical blur setting, and many are designed only for all-in-focus inputs. We present TextBokeh, a language-guided depth-of-field editor conditioned on a focal-object description and a relative defocus level. We train one model to synthesize bokeh from all-in-focus images and refocus already-defocused images at a requested defocus level. We construct its supervision by choosing the Chebyshev center of the target's estimated disparity interval as the focal plane, then holding this plane fixed while solving for the range of f-numbers that keeps the target sharp, produces visible off-target defocus, limits excessive blur, and respects the source lens range. Sampling this interval yields an ordered, scene-adaptive stack of training targets. Across the retained aperture levels, TextBokeh achieves substantially higher full-reference and perceptual fidelity than level-agnostic general-purpose editors on both AIF-to-Bokeh synthesis and cross-target refocusing. We will publicly release the trained LoRA adapter weights, the complete data synthesis pipeline, and the evaluation resources upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Depth-of-Field Editing, Bokeh Synthesis, Image Refocusing, Language-Guided Image Editing, Computational Photography

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7a12OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission proposes an algorithm for depth-of-field synthesis in images using language guidance. The user provides an image alongside a text prompt specifying which object in the image should remain in focus, as well as a numerical defocus strength level. The authors propose to fine-tune a foundation model using a synthetic dataset of paired images, whose construction appears to be the main technical contribution of this work. Specifically, several foundation models are arranged into a pipeline to generate training data pairs for a large collection of all-in-focus images. Experiments show that the approach performs favorably compared to a reasonable selection of baselines.

Critical strengths

Being able to edit the depth of field in an image is a desirable capability in many image editing workflows, and numerous solutions have been proposed over the decades. Instead of relying on image-level manipulations, this work addresses the problem by allowing users to specify their desired results through text prompts. This is, of course, a very common paradigm these days. At least two other well-known approaches exist in this space, but they apparently do not provide source code, preventing direct comparisons. The results of this work are quite convincing. The evaluation suggests that the proposed approach advances the state of the art.

Critical weaknesses

I have two closely related concerns about this submission.

My first concern relates to the mathematical derivations in Sec. 3.1, which are listed among the main contributions in Sec. 1. The presented derivations are elementary analytical results that have long been known in the field. For example, the harmonic-mean relationship in Eq. 5 can be found in an optics-industry technical white paper (https://www.photigy.com/wp-content/uploads/2010/09/DOF-and-Bokeh-explained.pdf), where it is formulated in terms of metric depth rather than inverse depth. Similar arguments apply to most other parts of the corresponding subsection. I am concerned that the exposition in the paper presents these elementary analytical results as substantial algorithmic contributions. I would strongly suggest more clearly delineating the original contributions in the exposition.

My second, related concern is that the manuscript does not adequately position itself with respect to the extensive body of related work that has accumulated over decades in the field of depth-of-field synthesis. The paper refers to "early methods" (L102), but only cites works published since 2016, with most of the discussed works being less than three years old. For example, the thin-lens equations considered here are accompanied only by a reference from 2022. The original sources in this area should be cited to give proper credit. Starting points for a literature review could be the following two works:

  • Potmesil & Chakravarty, A Lens and Aperture Camera Model for Synthetic Image Generation, SIGGRAPH 1981.
  • Kraus & Strengert, Depth-of-Field Rendering by Pyramidal Image Processing, CGF 2007.

I strongly encourage the authors to more seriously consider and acknowledge the rich history of the field to avoid the perception of insufficient scholarly rigor and a lack of due credit to foundational contributions.

Questions for authors

I do not have any questions.

Reviewer_VC5xOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TextBokeh formulates depth-of-field editing as a language-conditioned transition between focus states. A user describes the object that should be sharp and selects an ordinal defocus level from 1 to 5; the same editor handles both all-in-focus (AIF) inputs and already-defocused inputs. The main technical contribution is the supervision construction rather than a new generative architecture. For each language-specified target, the authors estimate a mask and metric depth, place the focal disparity at the midpoint of the target's disparity interval, and derive a feasible f-number interval subject to target-interior sharpness, off-target defocus, and lens constraints. Five candidates are sampled in log f-number, snapped to standard aperture values, and used to render target-specific supervision with BokehMe. The resulting dataset contains 95,307 source photographs, 306,011 focal targets, 829,351 rendered states, and 1,829,353 editing pairs. A rank-64 LoRA adapter trains Qwen-Image-Edit-2511 for both AIF-to-Bokeh synthesis and cross-target refocusing. The synthetic evaluation uses separate Open Images sources and reports strong PSNR, SSIM, LPIPS, and DISTS results against five general-purpose editors. Additional experiments consider 96 real-defocus photographs through GenFocus-generated AIF proxies, qualitative refocusing on real camera inputs, and same-target strength adjustment without explicit training pairs for that transition.

Critical strengths
  1. A useful interface with a coherent unified scope

Object-specific language control is a meaningful step beyond coarse instructions such as focusing on the foreground or background. TextBokeh lets the user name a small object part, a particular person, or a distant object, while removing depth maps, masks, focal distances, and aperture values from the inference interface. Supporting both AIF synthesis and transitions from defocused inputs in one model makes the interface considerably more useful than a bokeh-only generator. The same-target strength adjustment is also an interesting emergent transition, since those pairs are excluded from training. The qualitative results support that this is more than a generic blur prompt. The paired statue-eye examples (bokeh_gen/sample_0401 and sample_0402) are particularly effective: the outputs favor the requested left or right eye while retaining stone texture. sample_0450 keeps the small license-plate text legible at L5, and the level-change pair sample_0642/sample_0643 changes blur in both directions while preserving the racing wheel and framing. Across the supplied examples, TextBokeh's most consistent advantage is layout and content preservation; it largely avoids the cropping, recoloring, and object replacement seen in several general-purpose editors.

  1. A principled supervision idea executed at substantial scale

The minimax focal plane is mathematically elementary, but it is well matched to the problem: it maximizes the allowable blur under the stated target-interior constraint, and the feasible aperture interval then excludes states that violate sharpness, defocus, or lens limits. This is more principled than assigning one arbitrary blur range to every target. The execution is also substantial. The authors manually curated 5,000 AIF seed photographs, constructed a large automated annotation and rendering pipeline, trained on 1.83M pairs, and supplied detailed optical settings, data distributions, per-level counts, exact prompts, and qualitative results. The explicit disclosure of oracle diagnostics and protocol limitations adds to the paper's credibility. I see a meaningful research and engineering contribution here, not merely a thin wrapper around an existing editor.

Critical weaknesses
  1. The evaluation does not independently validate the proposed supervision or its advantage over dedicated methods

The mathematical guarantee is narrower than the natural-language idea of “keeping the target sharp and producing visible off-target defocus.” It bounds the rendered blur by one pixel only on an eroded interior, not on the complete target or its boundary ring. Moreover, both the target interval and the off-target statistic use extrema of estimated disparities. A single erroneous depth value can therefore shift the focal plane or the feasible aperture interval. In particular, the minimum off-target-defocus constraint is imposed on the maximum disparity deviation outside the target. It ensures that at least one extremal off-target location reaches c_min; it does not ensure that a visible background region, or any substantial fraction of off-target pixels, becomes defocused. An area-weighted analysis on the existing masks, depths, and retained f-numbers would test this directly without retraining. At present, the formulation proves a precise renderer-level property, but not necessarily the perceptual property that motivates the method. The main experiments also construct both training targets and test references with BokehMe and the same type of estimated masks and geometry. This is a valid way to measure fidelity to the intended synthesis pipeline, and the separate Open Images sources avoid train/test leakage. However, it cannot reveal shared mask, depth, or renderer errors: the editor can score well by reproducing the same incorrect focal plane or boundary artifacts present in a reference. Against this background, comparisons solely with general-purpose editors leave the central attribution unresolved. The results show effective specialization to the authors' references, but not how much the minimax plane and feasible interval contribute over simpler supervision, nor whether the model outperforms dedicated depth-of-field methods in image quality. The unavailability of BokehFlow and CamEdit is a reasonable practical constraint, and GenFocus has a different interface. Nevertheless, a native-controls GenFocus diagnostic would still be informative: the evaluation pipeline already computes geometric controls, and GenFocus is already used to create the real-defocus proxies. Such a comparison would separate the benefit of the language interface from the quality of the underlying editing operation. The comparison also mixes optical quality with instruction compliance. For example, Hunyuan's disclosed native rewrite changes the requested blue backdrop into the white sports car in bokeh_gen/sample_0440. Scoring that output against references for the original target is meaningful for overall task success, but it does not isolate optical bokeh quality. TextBokeh's superior scene preservation remains valuable; it is simply a different advantage from matching a particular defocus strength.

  1. Ordered 1–5 control is only indirectly tested, and the real-image evidence addresses a narrower task

Only 36,408 of 306,011 training targets retain all five levels, while 51% retain just one or two. Evaluation is restricted to retained labels, covering 79.7% of the nominal AIF grid and 71.9% of the nominal refocusing grid. The paper needs to explain what output is expected when a user requests a level that would have been pruned during supervision. It would also be useful to understand why the evaluation stacks are denser than the training stacks: the AIF test set averages 3.99 retained levels per case, compared with 2.71 in training, but the case-selection rule is not described. More fundamentally, exact-level PSNR/LPIPS is not by itself a direct test of level control. References from the same scene share almost all content, so a relatively level-insensitive prediction could still achieve good fidelity. The per-level tables also use different case subsets. A within-scene cross-scoring analysis—comparing every predicted level with every reference level—would directly test whether the requested level is the best match and whether predictions respect the known blur ordering. Much of this can be computed from outputs and references that already exist. The supplementary outputs are nevertheless encouraging about broad ordering. In a rough global gradient-energy diagnostic, only 7 of 344 adjacent available-level transitions move in the direction opposite to increased blur, and the visible reversals are mild. This is not a target-aware or strict proof of monotonicity, but it is positive evidence. The more important issue is calibration and target response: refocus/sample_0005 L5 under-blurs the non-target foreground car, whereas bokeh_gen/sample_0067 L4 over-blurs the jet nozzles relative to their references. In refocus/sample_0068, TextBokeh heavily blurs the white foreground car but leaves the requested red PORSCHE sign blurred as well. These cases do not negate the averages, but they show why target sharpness, non-target blur, and level calibration should be evaluated separately. The real-defocus experiment has a related limitation. Table 3 measures re-rendering from an AIF proxy produced by GenFocus, with the original photograph as reference and an oracle choosing among five TextBokeh outputs. The authors explicitly and appropriately state that Best-Pred is neither deployed level selection nor a test of arbitrary real-defocused-input refocusing. Consequently, this table supports candidate-set fidelity, but not prescribed-level accuracy or quantitative real-world refocusing. Direct refocusing on real camera inputs remains qualitative. The supplementary real-defocus viewer displays L5 outputs, but the real references have no demonstrated L5 calibration; these visual comparisons should likewise not be read as level-accuracy evidence. A simple input-identity reference would help calibrate all of these results. As a sanity check, I computed RGB PSNR on the supplied 768-pixel supplementary JPEGs. These numbers are not directly comparable to the paper's 1024-pixel tables, but TextBokeh's L5 outputs improve over the AIF proxy by approximately 1.42 dB on average across the 31 real-defocus cases, and the level-change outputs improve over copy-input by approximately 1.34 dB across the 38 supplied transitions, which cover 21 scenes/prompts. This is encouraging evidence that the model is doing meaningful work. The gains are not universal, however: 26/31 and 26/38 improve in the respective comparisons. Reporting the official identity and fixed-level results would therefore add useful context without requiring another training run.

Questions for authors
  1. What evidence isolates the benefit of the proposed focal-plane and aperture-interval construction from large-scale specialized fine-tuning? A controlled comparison on a subset with a simpler rule, such as median target disparity with conventional log-spaced apertures, would speak directly to the main contribution. On the existing supervision, what fraction of off-target area actually exceeds c_min, and what blur distribution occurs on the target boundary ring?
  2. How should the model handle a request for a level pruned during aperture-state construction? Can you report within-scene diagonal-versus-off-diagonal fidelity and blur ordering across levels, ideally including the pre-deduplication candidate references? Please also clarify the evaluation-case selection rule and why the test stacks are substantially denser than the training stacks.
  3. Can you report input-identity and fixed-level results for the existing real-defocus and same-target strength-adjustment outputs? These would distinguish genuine transformation from a strong input proxy and would complement the oracle Best-Pred analysis. Target sharpness and non-target blur would be especially informative alongside the full-image metrics.
Reviewer_ky1hOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TextBokeh, a language-guided framework for object-level focus and defocus control. It introduces a target-preserving aperture strategy to generate adaptive blur levels. The authors construct a dataset with 95K images and 1.83M training pairs, fine-tuning Qwen-Image-Edit-2511 with LoRA. Experiments demonstrate promising results in bokeh synthesis and cross-target refocusing.

Critical strengths
  1. This paper proposes a well-motivated and interpretable target-preserving aperture supervision strategy that maintains object sharpness across a non-negligible depth range. It provides scene-adaptive supervision rather than relying on fixed blur-strength settings.
  2. TextBokeh presents a unified framework for two practical image editing tasks, integrating AIF-to-bokeh synthesis and cross-target refocusing into a single language-guided model.
Critical weaknesses
  1. This paper uses BokehMe to generate depth-of-field supervision images during training, while the synthetic reference images used for evaluation are also primarily rendered with BokehMe. This may lead the model to mainly learn the rendering characteristics of BokehMe rather than general physical optical bokeh effects. More extensive experiments are needed to demonstrate TextBokeh's generalization to real-world camera imaging conditions.
  2. This paper mainly compares TextBokeh with general-purpose image editing models, such as Qwen, FLUX, and Hunyuan, but lacks comparisons with existing specialized depth-of-field editing methods.
  3. An important aspect of bokeh rendering is whether subject boundaries remain sharp and well-preserved. More visual results, particularly close-up comparisons, are needed to evaluate this aspect.
  4. The five-level supervision data are highly imbalanced, and the real-world evaluation relies on oracle-based selection. Therefore, the accuracy and reliability of the requested defocus-level control have not been sufficiently demonstrated.
Questions for authors
  1. The model does not use depth maps during inference, with depth information only involved in training data construction. Therefore, it remains unclear whether the model truly learns the underlying principles of depth of field or merely generates bokeh-like effects based on semantic and visual appearance.

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

Related papers

Discussion (0)

Sign in to comment.