acceptodds
Under review as a conference paper at ICLR 2027

Inference-time Trajectory Correction for Structure-Preserving Manga Image Editing

Abstract

We present a lightweight, training-free trajectory correction method that adapts a pretrained image editing model to each input manga image using only the input itself. Despite recent progress in pretrained image editing, such models often underperform on manga because they are trained predominantly on natural-image data, while re-training or fine-tuning them on manga is costly and raises copyright concerns. Many manga image editing tasks encountered in practice are structure-preserving, requiring local details to be modified while the input's global composition is retained. To support this common editing setting, our method corrects the early editing trajectory by anchoring it to an empty-prompt reconstruction trajectory. Experiments on text removal show consistent improvements over direct inference with both Qwen Image Edit 2509 and FLUX.1 Kontext, and an additional application to screentone synthesis shows that the method also better preserves the input composition. With FLUX.1 Kontext on an RTX A6000, the method incurs 11% runtime overhead and 0.1% peak-memory overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
manga, image editing, rectified flow, flow matching, inference-time adaptation

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_x5jUOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an inference-time trajectory correction method for structure-preserving manga image editing, which aims to adapt pretrained image editing models to manga images without additional training or architectural modifications. The authors point out that existing pretrained image editing models often struggle to preserve the original image structure during local editing due to the domain gap between natural images and manga images. Motivated by this observation, the proposed method aims to preserve the global composition of the input image while allowing the model to perform the intended local edits by correcting the early stages of the generation trajectory. Specifically, the authors leverage empty-prompt velocity predictions to construct an input-specific reconstruction target and interpolate the current latent state toward this target, while retaining the edit-prompt velocity for the subsequent sampling update. The authors further provide a mathematical analysis to explain why empty-prompt-based correction can preserve the editing response more effectively than edit-prompt-based correction. Experiments on manga text removal and screentone synthesis demonstrate improvements in structure preservation across different pretrained image editing models, with relatively low computational overhead.

Critical strengths
  1. In terms of empirical performance, the proposed method achieves the best or highly competitive results on most evaluation metrics, demonstrating its effectiveness in preserving image structure while performing the intended edits. The consistent improvements across two pretrained image editing models further support the practical value of the approach.
  2. The proposed method is simple, intuitive, and easy to understand and implement, requiring no additional training or architectural modifications. Moreover, Section 4.2, Why the Empty Prompt Helps, provides a clear mathematical explanation of the key design choice. In particular, it shows why using the empty-prompt velocity to construct the correction target avoids explicitly weakening the editing response, making the proposed approach more convincing and well-motivated.
Critical weaknesses
  1. Table 1 shows that the pretrained model achieves a low PSNR of 11.84 dB under empty-prompt reconstruction, whereas the proposed method improves it to 22.27 dB. Given that the correction relies on empty-prompt velocity predictions from the same model, it remains unclear how these predictions enable such a substantial improvement despite the poor initial reconstruction performance. Further analysis of the correction's effect on the generation trajectory would help clarify this result.
  2. In Table 2, the proposed method and several competing approaches, including FlowChef and FlowEdit, are evaluated using different model backbones. Although the authors explain that these methods are not directly compatible with instruction-based image editing models, the backbone differences make it difficult to attribute the observed performance gaps solely to the proposed method. A more controlled comparison would strengthen the experimental evidence.
Questions for authors
  1. Table 1 shows a substantial gap between the baseline (11.84 dB) and the proposed method (22.27 dB) under empty-prompt reconstruction. The authors should provide further experiments to explain how the proposed correction achieves such a large improvement despite the poor reconstruction quality of the original empty-prompt model. In particular, analyzing the reconstruction error and trajectory behavior before and after correction would help clarify the role and reliability of the empty-prompt velocity.
  2. The authors should conduct additional comparisons using the same model backbone whenever technically feasible. This would help isolate the contribution of the proposed trajectory correction from differences in model architectures and pretrained capabilities, providing stronger evidence of its effectiveness relative to existing methods.
5 comments
AuthorsOpenReview ↗

[Comment 1 of 1] Title: Response to Reviewer x5jU

Thank you for the helpful comments. In short: (Q1) a trajectory analysis shows that each early corrected step reduces the gap between the model's predicted clean image and the input, which explains the gain from 11.84 dB to 22.27 dB; (Q2) on the same image editing model as ours, our method is the only training-free method that improves all five metrics. Details follow, with pointers to the revised paper.

Q1: Why does reconstruction improve from 11.84 dB to 22.27 dB?

Notation: at each sampling step, Z_t is the current latent, X_in is the input image (in latent space), u is the velocity predicted with the empty prompt, and Δ is the remaining time to the end of sampling, so Ẑ = Z_t + Δu is the clean image that the model currently predicts. Our correction target (Eq. 5) can be rewritten as Z* = Z_t + (X_in − Ẑ). The target therefore shifts the current latent by the gap between the input and this prediction, so that following u from the target would end exactly at the input. In other words, the content of the target comes from the input, and u is used only to convert the input image into the corresponding state at the current noise level. At every corrected step, we do not replace the current latent with this target but move it only a fraction α = 0.01 toward the target (Eq. 6). Each corrected step thus reduces the gap between the prediction and the input. The baseline in Table 1 instead follows the model's velocity without such a correction and drifts away from the input, which is consistent with its 11.84 dB.

We verified this by tracking the sampling trajectories of 100 test images with Qwen Image Edit 2509 under the empty prompt, measuring at every step the residual between the predicted clean image and the input, ‖Ẑ − X_in‖/‖X_in‖ (Sec. 5.4, Fig. 3):

  • Both methods start at a residual of 0.45. The baseline residual grows to 0.58 by the output, whereas ours drops to 0.13 within the three corrected steps and stays below 0.18 until the output. Proposition 1 alone predicts that each correction with α = 0.01 reduces the residual by only 1%. The actual reduction is much larger because the model, evaluated at the slightly corrected latent, predicts a new velocity whose predicted clean image is closer to the input.

Q2: Comparison on the same backbone

FlowChef and FlowEdit were designed for text-to-image models, which is why the submission ran them on FLUX.1 [dev] and Qwen Image 2512 rather than on an instruction-based image editing model. For a controlled comparison, we adapted both to run on the same image editing model as ours, Qwen Image Edit 2509 (denoted "edit mode" in the revised paper), and added GRAG (CVPR 2026 Findings), a training-free method designed for instruction-based image editing models (Table 2, Supp. B.5). All methods use the baseline's instruction ("Remove text. Keep everything else unchanged."), classifier-free guidance scale, number of steps, and random seed, and each method uses its default setting.

Method (all on Qwen Image Edit 2509)PSNR ↑SSIM ↑LPIPS ↓Text PSNR ↑Text SSIM ↑
Baseline19.6272.767.9219.0760.77
FlowChef, edit mode21.5780.686.7418.6961.99
FlowEdit, edit mode21.0589.8810.289.7525.41
GRAG23.6486.515.1417.9762.03
Ours, α=0.01 (fixed)22.0780.616.0020.0464.15
Ours, α=0.03 (reference only, Supp. Table 5)24.3987.435.1020.7664.49

The Text columns are computed inside the ground-truth text mask. With α = 0.01, which we fix for all backbones and tasks, ours is the only method that improves all five metrics over the baseline. The other methods raise the full-image scores mainly by keeping more of the input, including the text that should be removed. GRAG thus obtains higher full-image scores than ours but fails to remove much of the text. Its text-region PSNR (17.97) is even lower than that of the baseline (19.07), whereas ours raises it to 20.04. FlowChef in edit mode also falls below the baseline in text-region PSNR, and FlowEdit in edit mode leaves most of the image, text included, unchanged, which explains its high SSIM. For reference, our hyperparameter study (Supp. Table 5) shows that α = 0.03 further improves all five metrics on this backbone and exceeds GRAG on all of them (last row).

(The baseline LPIPS in Table 2 has been corrected from 7.16 to 7.92; see the general response.)

We hope these responses address your concerns. If anything remains unclear, we would welcome further discussion.

Thank you for the response. The additional experiments address my main concerns. However, the method appears sensitive to the choice of hyperparameters. Therefore, I will maintain my rating of 3.

AuthorsOpenReview ↗

Response to Reviewer x5jU (follow-up)

Thank you for confirming that the additional experiments address your main concerns. We would like to clarify the remaining point on hyperparameter sensitivity, because the results in the paper show that the proposed method does not require tuning.

A single fixed setting, not tuned on test data

All the main results of the paper use the same setting, correcting the first M = 3 sampling steps with strength α = 0.01, which we did not tune on any test set and applied unchanged to the newer backbone FLUX.2 32B. With this one setting, the proposed correction improves all five text-removal metrics over the baseline for Qwen Image Edit 2509, FLUX.1 Kontext, and FLUX.2 32B (Table 2). It also improves all nine metrics of document highlighter removal (Supp. D). More generally, the paper uses this same setting for every task, without per-task tuning (Supp. B.3).

Not sensitive within a practical range

The supplementary material has included a hyperparameter study since the submission (Supp. B.3, Table 5), and it shows that the method is not sensitive to its hyperparameters within a practical range. Specifically, on Qwen Image Edit 2509, every tested value of M from 1 to 7 and of α from 0.01 to 0.05 improves all five text-removal metrics over the baseline, so no careful tuning is needed. Table 5 also lists settings that we deliberately chose far outside this range, such as α = 0.001 and α = 0.1, which are one tenth and ten times the default. They are not candidate operating points. Instead, we include them to illustrate under-correction (too weak to correct the trajectory) and over-correction (the output stays too close to the input and text remains), which explains the role of each hyperparameter. The metrics that degrade in these rows therefore reflect these intentionally extreme choices and do not indicate sensitivity within the practical range.

The fixed setting is enough in practice

In the comparison of training-free methods on the same backbone, Qwen Image Edit 2509 (the table in our first response and Table 2), ours with this fixed setting is the only training-free method that improves all five metrics over the baseline. In contrast, each of the other methods, at its default setting, falls below the baseline in at least one metric. Beyond this, the α = 0.03 row of our first response (Supp. Table 5) shows that the strength of the correction leaves room for further gains. These gains are optional. In general, the strength that gives the best balance between keeping the input and applying the edit naturally depends on the task, as for any method that balances the two. However, the key point is that the proposed method already improves over the baseline with its single fixed setting (e.g., in text removal and document highlighter removal above), and all the main results in the paper use this setting.

We hope this clarifies that the method works with one fixed setting across these models and tasks without careful tuning, and that adjusting its strength is an option for further gains rather than a requirement. If any question or concern remains, we would be glad to respond during the remaining discussion period.

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

This paper proposes a training free method for Manga image editing, focusing on text removal, and minor appearance changes. They build on pretrained image editing models (Qwen-image-edit). Their key observation is that Manga images are out-of-distribution for such models, and therefore editing such images with these models results in unwanted changes. Assuming that providing "empty prompt" to the model should reconstruct the input, they correct the denoising trajectory by replacing the clean image prediction with empty prompt of the model, with the input image, and then slightly moving the noisy latent at timestep t towards this corrected trajectory. By doing this they encourage structure preservation and successful edits.

Critical strengths
  • The presented idea is simple and smart

  • The presented results are nice and show the effectiveness of the model

Critical weaknesses
  • The setting that the work targets in too limited. I don't undersand what is the exact purpose of focusing solely on text removal in Manga images. I think it is necessary to see more diverse editing tasks, even if they are only minor appearance editing tasks

  • I checked the example provided in the supplementary with Flux 2 and Flux 3 Klein, and both models were able to successfully remove the text while preserving the structure of the image. I think there needs to be a discussion about newer models such as Flux 2 (Not Flux 3 since its release date was after the ICLR deadline

  • Given the limited setting, I do not fully agree with the motivation of the proposed method were training LoRAs need a lot of data that requires copyrighted material. Given the model focuses only on text removal, I think even curating a small synthetic data with text added/removed can successfully perform this type of edit

  • While the proposed method works, the main assumption that empty prompt should reconstruct the image is not clear to me. I think this needs further motivation. Explicitly prior work has already explored this and shows that this is not the case (Continuous Control of Editing Models via Adaptive-Origin Guidance, Wolf et al.)

Questions for authors

See above

5 comments
AuthorsOpenReview ↗

[Comment 1 of 2] Title: Response to Reviewer hKVu (1/2)

Thank you for the helpful comments. In short: (W1) besides text removal, the paper covers screentone synthesis, now with line-art structure metrics, and we added document highlighter removal; (W2) on FLUX.2 32B, the proposed correction improves all five text-removal metrics with statistically significant differences, mostly on images where the baseline deviates from the input; (W3) even a small synthetic dataset would be made from manga images, which cannot be used for training without permission, and would have to be prepared for each task; (W4) the method uses the empty prompt only to compute the state from which the input would be reached, so it does not assume that the empty prompt reconstructs the input. Details follow.

W1: Limited setting

The paper is not limited to text removal: since the submission, the supplementary material has reported screentone synthesis, which converts line art into a finished manga image and thus adds content. We chose text removal as the main task because it is important in practice (manga localization) and allows exact ground truth and text masks. In screentone synthesis, the input line art is extracted from a finished manga image with a line-extraction model (Manga Line Extraction). In the revision, we also measure structure preservation directly. We apply the same model to each output and compare the extracted line art with the input line art. With Qwen Image Edit 2511, the image editing model used in the paper, and with FLUX.2 32B, all three line-art metrics improve (Supp. C.3, Table 9). The SSIM gain with Qwen Image Edit 2511 is statistically significant.

Model (baseline → ours)Line-art PSNR ↑Line-art SSIM ↑Line-art LPIPS ↓
Qwen Image Edit 251114.87 → 15.4680.17 → 83.0014.75 → 13.03
FLUX.2 32B17.99 → 18.6089.22 → 90.277.54 → 7.23

As an appearance edit beyond manga, removing highlighter marks from document images also improves with the fixed setting, which shows that the same training-free correction carries over to other tasks (see our response to Reviewer G5ee and Supp. D).

W2: Newer models such as FLUX.2

Thank you for testing our example; on that image, both the baseline and ours remove the text. We evaluated FLUX.2 32B, the open-weight FLUX.2 model released before the deadline, on the full text-removal benchmark of 3,204 image pairs, using the same hyperparameters as for Qwen Image Edit 2509 (M = 3, α = 0.01) without any tuning for this model (Table 2, Supp. B.6):

Method (FLUX.2 32B)PSNR ↑SSIM ↑LPIPS ↓Text PSNR ↑Text SSIM ↑
Baseline27.8695.133.5220.3171.73
Ours28.0595.723.3320.5071.87
  • All five text-removal metrics improve with statistically significant differences. When we split the images into four equal groups by the baseline's full-image SSIM, the gain is largest in the group where the baseline deviates most from the input, while images that the baseline already handles well, such as the example in our supplementary code that you tested, stay nearly unchanged. The table below shows the change from the baseline to ours (Δ) in each group (Supp. Table 7).
Quartile by baseline SSIMBaseline SSIMΔSSIM ↑ΔPSNR ↑ΔLPIPS ↓
Q1 (lowest)90.37+1.48+0.43−0.61
Q295.27+0.52+0.23−0.13
Q396.84+0.26+0.10−0.04
Q4 (highest)98.04+0.09+0.02+0.02
  • In screentone synthesis, where the composition is more easily altered, FID improves from 42.31 to 39.82, CMMD from 0.3194 to 0.2995, and line-art SSIM from 89.22 to 90.27 (Supp. C.3).
AuthorsOpenReview ↗

[Comment 2 of 2] Title: Response to Reviewer hKVu (2/2)

W3: Training a LoRA on a small synthetic dataset

We apologize that our motivation was unclear, and we have revised Sec. 1. We agree that a LoRA trained on a small synthetic dataset may work for text removal. However, such a dataset is still made from manga images, for example by adding text to text-free manga pages, and in practice manga images cannot be used for training without permission from the rights holders, even in small numbers. In addition, such data would have to be prepared for each editing task. Our method needs neither manga training data nor task-specific data. It handles text removal, screentone synthesis, and document highlighter removal with the same fixed hyperparameter setting by changing only the instruction. Because it only modifies the sampler, it also applies to new backbones such as FLUX.2 32B without retraining.

W4: Whether the method assumes that the empty prompt reconstructs the input

First, a clarification: our method does not assume that the empty prompt reconstructs the input. As your summary describes, at each of the first few steps, the correction replaces the clean image predicted with the empty prompt by the input image and moves the latent slightly toward the resulting state, which keeps the global structure of the input. The content of this target comes from the input itself, and the empty prompt is used only to convert the input image into the corresponding state at the current noise level. The model thus does not need to reconstruct the input under the empty prompt. Indeed, it does not (11.84 dB in Table 1), consistent with the observation of Wolf et al. that the null instruction acts as an arbitrary edit. Yet the correction built from it brings the reconstruction to 22.27 dB, and the trajectory analysis shows how (see our response to Reviewer x5jU and Sec. 5.4). To make this clearer, we revised the wording of the abstract and Sec. 3.2. We use the empty prompt rather than the edit prompt for the target because a target built with the edit prompt weakens the response to the edit instruction, whereas the empty prompt keeps it unchanged (Proposition 2 in Sec. 4.2, Supp. B.1). AdaOr handles this behavior of the null instruction by training an identity instruction token, whereas our method needs no training because it does not require the empty prompt to reconstruct the input (Sec. 2.1).

We hope these responses address your concerns. If anything remains unclear, we would welcome further discussion.

Reviewer_hKVuOpenReview ↗

Response to rebuttal

Thanks for the rebuttal. Some of my concerns have been addressed. I appreciate the simplicity of the method and I think both the observation and the proposed fix are sound and interesting. But my concerns about the motivation of the work and the scope of the results presented still remain. On the scope of the work, I think it would be much stronger to follow the evaluation protocol of prior art such as FlowEdit and FlowChef and compare on general (only structure preserving) edits for example on Pie-bench. I guess this method would also work in these settings, but providing such results are essential to prove that the proposed method is really comparable with SOTA (FlowEdit, FlowChef, Flux2, etc). On the motivation, I'm not sure if this is a fair comment but I still think that the problem is solvable simply with scale or finetuning (The authors mention copyright restrictions, but my original point was that even finetuning on a small data of non-copyrighted (even non-manga) data would perform better. That is why I still insist on evaluating on more general editing settings. And for scale We already see on the newer results that Flux 2 does close the gap between this method and the baseline and the numbers are really close. There is also no qualitative results on Flux2.dev to see whether the baseline's errors are even noticeable. With all this being said, I still like the method, and I would not argue too strongly for rejection and I raise my score to weak reject.

Reviewer_G5eeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a lightweight, training-free trajectory correction for instruction-based image editing models that preserves global composition during local manga edits. The method mainly focuses on moving the estimated endpoint toward the input without weakening the editing response during each correction step. Experiments with Qwen Image Edit and FLUX indicate improved text removal over direct inference.

Critical strengths
  1. Good writing, very easy to read and follow; motivations are explained very clearly.
  2. The method is simple and effective, and the related theoretical analysis is clear.
Critical weaknesses
  1. Although effective, this paper mainly focuses on manga image editing, which limits its generalizability.

  2. My main concern is about the experiments. Considering that instruction-based editing models like Qwen-Image Editing usually have much more prior, I think that better performance than inverse-based methods like FlowEdit can not support the effectiveness of the proposed method. However, although there are many inversion-based editing baselines, only FlowChef is an instruction-based editing method.

Questions for authors

Besides FlowChef, could the authors discuss three more instruction-based editing methods, including their mechanisms and performance on this task?

2 comments
AuthorsOpenReview ↗

[Comment 1 of 2] Title: Response to Reviewer G5ee (1/2)

Thank you for the helpful comments and for suggesting a comparison with instruction-based methods. In short: (W1) we added a non-manga task, highlighter removal from document images, where the proposed correction improves all nine metrics without any training; (W2/Q) we compared three training-free methods running on the same instruction-based image editing model as ours, and ours is the only one that improves all five metrics. Details follow.

W1: Focus on manga

Our method changes only the sampler of a pretrained image editing model and contains nothing specific to manga. We focus on manga because it is far from the training data of image editing models, so their outputs often lose the input composition (Fig. 1), and because copyright restricts training on manga, which calls for a training-free method. The method itself, however, is technically applicable to other images that image editing models fail to edit faithfully. To test another out-of-distribution domain, we applied the method, with the fixed setting and only the instruction changed, to removing highlighter marks from document images. We drew synthetic yellow highlighter strokes on 1,000 real pages from DocLayNet (financial reports, scientific articles, patents, laws, tenders, and manuals), and the original page serves as the ground truth. All nine metrics in the table below improve with statistically significant differences, and LPIPS is lower on 92% of the pages (Sec. 5.5, Supp. D).

Region (baseline → ours)PSNR ↑SSIM ↑LPIPS ↓
Full page16.08 → 16.3568.19 → 69.7213.64 → 12.51
Inside the highlighted region12.29 → 12.4828.81 → 31.652.68 → 2.43
Outside the highlighted region16.39 → 16.6869.85 → 71.3313.13 → 12.03

This experiment is meant to show the generality of the proposed correction, i.e., that it improves an existing image editing model on another task without any training, rather than to claim state-of-the-art document editing.

AuthorsOpenReview ↗

[Comment 2 of 2] Title: Response to Reviewer G5ee (2/2)

W2 / Q: Comparison with instruction-based methods

First, a clarification: FlowChef is not an instruction-based method; it steers the sampling of a text-to-image model toward an objective. We agree that outperforming methods built on text-to-image models is not sufficient, because instruction-based image editing models have much stronger priors. We therefore ran three training-free methods, GRAG, FlowEdit, and FlowChef, on the same instruction-based image editing model as ours, Qwen Image Edit 2509, with the same instruction, guidance scale, number of steps, and seed (Table 2, Supp. B.5). FlowEdit and FlowChef were designed for text-to-image models, so we adapted them to this image editing model. We refer to the adapted versions as "edit mode", as in the revised paper. We also discuss AdaOr, which requires training.

  • GRAG (CVPR 2026 Findings): a training-free guidance for instruction-based image editing models. In each attention layer, it splits the keys of the input-image tokens into their mean and the deviations from it, and scales the deviations by a fixed factor (default 1.1) to trade off fidelity to the input against edit strength. At its default, it raises the full-image scores (PSNR 23.64) but also keeps the text, so the text-region PSNR falls below the baseline (19.07→17.97).
  • FlowEdit in edit mode: FlowEdit edits without inversion by integrating the difference between a target velocity and a source velocity at noisy versions of the input. We use the velocity of the image editing model with the empty instruction as the source and with the edit instruction as the target. At its default number of editing steps, it removes little text (text-region PSNR 9.75).
  • FlowChef in edit mode: pulls the clean image predicted with the edit instruction toward the input in the early steps. This is exactly the edit-prompt variant of our method analyzed in Proposition 2, which weakens the response to the instruction (text-region PSNR 18.69).
  • AdaOr (SIGGRAPH 2026): an instruction-based editing method that trains an identity instruction token (a LoRA for Qwen Image Edit) and interpolates between the identity and null predictions to control the edit strength. Because it requires training the image editing model, it falls outside our training-free setting, and we did not run it. AdaOr adjusts the strength of the whole edit, so keeping more of the input also means a weaker edit. Our correction instead keeps the overall layout of the input without weakening the requested edit, such as text removal (Proposition 2 in Sec. 4.2).

With α = 0.01, which we fix for all backbones and tasks, ours is the only one that improves all five metrics over the unmodified image editing model (the full results are in the table of our response to Reviewer x5jU). GRAG scores higher than ours on the full-image metrics but fails to remove much of the text, and its text-region PSNR (17.97) is even lower than that of the baseline (19.07). For reference, our hyperparameter study (Supp. Table 5) shows that α = 0.03 further improves all five metrics and exceeds GRAG on all of them.

We hope these responses address your concerns. If anything remains unclear, we would welcome further discussion.

Comments on the paper

2 comments
AuthorsOpenReview ↗

[Comment 1 of 2] Title: Summary of concerns and responses (1/2)

We thank all reviewers for their constructive feedback. To address the comments, we conducted five additional experiments and revised the paper. Changes are shown in blue. The main text grows by less than one page, and the details are in the supplementary material ("Supp." below). Throughout, the compared methods use their default settings, and our method keeps the fixed hyperparameter setting used in the paper: the proposed correction is applied to the first M=3 sampling steps with strength α=0.01. No hyperparameter was tuned on the test data. The α=0.03 results quoted in our responses come from our hyperparameter study (Supp. Table 5) and are given only for reference. In the text-removal results, "full-image" metrics are computed over the whole image and "text-region" metrics only inside the ground-truth text mask, i.e., where the text must be removed. The table below lists each concern raised in the reviews, the reviewers who raised it, our response, and where it is addressed in the revised paper. Each individual response is self-contained and refers to the responses to other reviewers where the concerns overlap.

Concerns and our responses (1–4 of 6)

ConcernOur responseRaised byLocation
1. The experiments focus on text removal from manga, which limits the generality of the method.The method changes only the sampler of a pretrained image editing model and contains nothing specific to manga. Besides text removal, the paper covers screentone synthesis, which adds content to line art. We now also measure structure preservation in this task by comparing the line art extracted from each output with the input line art. Line-art SSIM improves from 80.17 to 83.00 with Qwen Image Edit 2511. We also added a non-manga task, highlighter removal from 1,000 real document pages, where all nine metrics (PSNR, SSIM, and LPIPS over the page, inside the highlighted region, and outside it) improve with the same fixed hyperparameter setting.G5ee W1, hKVu W1Sec. 5.5, Supp. C.3, D
2. FlowChef and FlowEdit run on different backbones from ours, so the gains may come from the stronger prior of the instruction-based image editing model rather than from the method. Comparisons with instruction-based methods on the same backbone are needed.We ran three training-free methods on Qwen Image Edit 2509, the same instruction-based image editing model as ours. These are GRAG (CVPR 2026 Findings), which is designed for such models, and FlowChef and FlowEdit, which we adapted to run on it. Ours is the only method that improves all five text-removal metrics over the baseline. GRAG obtains higher full-image scores but fails to remove much of the text, and its text-region PSNR falls below that of the baseline. Secs. 1 and 2.1 now state how our method differs from GRAG and from AdaOr (SIGGRAPH 2026), which requires training, and limit our novelty claim to input-specific test-time adaptation.x5jU Q2, G5ee W2/QTable 2, Secs. 1, 2.1, and 5.5, Supp. B.5
3. It is unclear how the empty-prompt predictions of the same model raise the reconstruction PSNR from 11.84 dB to 22.27 dB (Table 1).The content of the correction target comes from the input rather than from the model's prediction, so each corrected step reduces the gap between the predicted clean image and the input. On 100 test images, this gap drops from 0.45 to 0.13 within the three corrected steps and stays below 0.18 until the output, whereas it grows to 0.58 with the baseline.x5jU Q1Sec. 5.4, Fig. 3
4. The method seems to assume that the empty prompt reconstructs the input, although prior work (AdaOr) shows that it does not.As a clarification, the method does not make this assumption. It uses the empty prompt only to convert the input image into the corresponding state at the current noise level, and the content of the correction target comes from the input (see concern 3). We use the empty prompt rather than the edit prompt because a target built with the edit prompt weakens the response to the edit instruction (Proposition 2). We revised the abstract and Sec. 3.2 to make this clear, and Sec. 2.1 states the difference from AdaOr, which handles this behavior of the empty prompt by training.hKVu W4Abstract, Secs. 2.1, 3.2, and 4.2
AuthorsOpenReview ↗

[Comment 2 of 2] Title: Summary of concerns and responses (2/2)

Concerns and our responses (5–6 of 6)

ConcernOur responseRaised byLocation
5. Newer image editing models such as FLUX.2 should be discussed. On the provided example, FLUX.2 already removes the text while preserving the structure.We evaluated FLUX.2 32B on the full text-removal benchmark (3,204 image pairs) with the same hyperparameters as for Qwen Image Edit 2509 and without any tuning. All five metrics improve with statistically significant differences. The gain comes mostly from images where the baseline output deviates from the input, while images that the baseline already handles well, such as the provided example, stay nearly unchanged. In screentone synthesis, FID improves from 42.31 to 39.82.hKVu W2Table 2, Sec. 5.5, Supp. B.6, C.3
6. The motivation for a training-free method is unclear, since a LoRA trained on a small synthetic dataset may suffice for text removal.We agree that such a LoRA may work for text removal. However, even a small synthetic dataset is made from manga images, which cannot be used for training without permission from the rights holders, and such data would be needed for each editing task. Our method needs no training data and handles text removal, screentone synthesis, and document highlighter removal with the same fixed hyperparameter setting by changing only the instruction. We revised Sec. 1 to make this motivation clear.hKVu W3Sec. 1

Corrections

  • The LPIPS of the Qwen Image Edit 2509 baseline in Table 2 and Supp. Tables 3–6 is 7.92, not 7.16 (a transcription error in the submission). Only this baseline value changes. Our result (6.00) and all other values are unchanged, and our method still improves LPIPS over the baseline, so the interpretation of the results is unaffected.
  • The screentone experiment mislabeled FLUX.2 Klein 9B as "Klein 2B", and the submission stated M=3 for all models. Klein 9B is distilled to sample in only four steps, so we applied the correction to its first step only (M=1). These are corrections of the description only. The reported numbers were obtained with this setting and are unchanged, so their interpretation is also unchanged.
  • AI use statement: we added that an AI agent assisted in designing, implementing, running, and analyzing the new experiments and in drafting the revised text.

We hope these responses and revisions address the reviewers' concerns. If anything remains unclear, we would welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.