While language-guided image manipulation has made remarkable progress, the challenge of how to instruct the manipulation process faithfully reflecting human intentions persists. An accurate and comprehensive description of a manipulation task using natural language is laborious and sometimes even impossible, primarily due to the inherent uncertainty and ambiguity present in linguistic expressions. Is it feasible to accomplish image manipulation without resorting to external cross-modal language information? If this possibility exists, the inherent modality gap would be effortlessly eliminated. In this paper, we propose a novel manipulation methodology, dubbed ImageBrush, that learns visual instructions for more accurate image editing. Our key idea is to employ a pair of transformation images as visual instructions, which not only precisely captures human intention but also facilitates accessibility in real-world scenarios. Capturing visual instructions is particularly challenging because it involves extracting the underlying intentions solely from visual demonstrations and then applying this operation to a new image. To address this challenge, we formulate visual instruction learning as a diffusion-based inpainting problem, where the contextual information is fully exploited through an iterative process of generation. A visual prompting encoder is carefully devised to enhance the model's capacity in uncovering human intent behind the visual instructions. Extensive experiments show that our method generates engaging manipulation results conforming to the transformations entailed in demonstrations. Moreover, our model exhibits robust generalization capabilities on various downstream tasks such as pose transfer, image translation and video inpainting.
Paper
Similar papers
Peer review
Summary
This paper proposed a novel scheme for image editing via exemplar-based. An image edit will be represented as an in-context instructions. To that end, the novel framework is trained to perform image editing as image in-painting task. Experiment results indicate the advantages of this scheme over typical text-guided image editing frameworks.
Strengths
**Originality/Novelty**: - At a high level, this paper proposes a novel protocol for image editing: visual prompting. I believe this is an area will attract interests of both academic and applications. - The designed framework, called ImageBrush, enable performing image manipulation in-context learning approach. Moreover, the interface design caters to user interest, which, although an add-on feature, proves beneficial for end-users. - Through experiments and visual example, the paper demonstrates the merits of proposed protocol, which can be more acurate and efficient for image editing tasks. - The authors did give efforts to try on multiple types of manipulation (e.g., seketon to fashion pose), instead of only standard image editing task. **Presentation/ Writing**: - Overall, the paper is well-organized and easy to follow. - Framework illustration is fairly good; qualitative examples depicts the merits of the method.
Weaknesses
**1. Some parts need to be revised to avoid over-claiming/ clearer** 1.1 The contribution mentioned in lines 72-73 overlaps with the Visual Prompting approach proposed in [1]. While I understand that this paper focuses on more specific tasks such as image editing and pose transfer, it would be advisable to use a more restrained tone. Instead of making strong claims about the original idea of visual prompting, the authors could emphasize the concepts of "progressive inpainting" or "prompting for additional tasks." 1.2 The "Related Work" section should include references to "Visual Prompting" or "In-context learning" to provide more complete survey of prior work. 1.3. Figure 3 is unclear and requires further clarification. Please explain this Figure in the rebuttal. **2. Flaws in experiments** 2.1. While the main contribution of this paper is an in-context learning scheme for image manipulation, I do not see any comparisons against existing in-context learning scheme (e.g., Visual Prompting [1] or Painter [2]). I am not against the fact that the authors demonstrate the effective against "state-of-the-art" methods for each specific tasks (e.g.: compare with InstructPix2Pix in image editing). I believe that it's crucial to compare against in-context learning methods as well. For example: Authors should compare image in-painting task with Visual Prompting [1]. 2.2. Lack of results for Interface design. While I second this contribution, there is no qualitative results for this module in the main paper. I have checked the Supplementary, only one example (edit of "putting a crown on the woman" is included). This example is not convincing itself, as I see the the crown in both approach is still not visually similar to the given crown in visual prompting. Additionally, I feel like these results can be generated by different noises. For example, given the same visual prompt (e.g., "Put a crown on the woman", different noises can also lead to different visually different crown). For that reason, while I second this module, I am not convinced that this module works in practice. 2.3. Unfair comparison in video inpainting task. Following [4] Video-FID (together with PSNR and SSIM) is used to evaluate the methods. Is it better to use Video-FID instead of FID? **3. Miscellaneous** 3.1 Line 124- 138: While cite Latent Diffusion model [5], the formula are in pixel-space. Should the citation update to DDPM instead to suit the following formular? 3.2 Line 243: Missing citation for the "CLIP Direction Similarity" mentioned. 3.3 Line 218: Is the dataset name correct? My understanding is the authors indeed use the [InstructPix2Pix's CLIP-filtered-dataset]((https://github.com/timothybrooks/instruct-pix2pix#generated-dataset). There might be a misunderstood, as in InstructPix2Pix, "LAION Improved Aesthetics 6.5+" is used as just a small part of generating the final dataset (as written in [this](https://github.com/timothybrooks/instruct-pix2pix#11-manually-write-a-dataset-of-instructions-and-captions)). **References** [1] Amir et al., *Visual Prompting via Image Inpainting*, NeurIPS 2022. [2] Xinlong et al., *A Generalist Painter for In-Context Visual Learning*, CVPR 2023. [3] InstructPix2Pix et al., *InstructPix2Pix: Learning to Follow Image Editing Instructions*, CVPR 2023. [4] Xueyan et al., *Progressive Temporal Feature Alignment Network for Video Inpainting*, CVPR 2021. [5] Robin et al., *High-Resolution Image Synthesis with Latent Diffusion Models*, CVPR 2022.
Questions
My initial rating is "Borderline accept" as I believe that while this paper is well-motivated, there are several revisions that need to be made. In addition to the questions raised in the Weakness section, I also have some technical inquiries: - Is the output size of all images 256x256? (which is smaller than other methods typically using 512x512). - Was the model fine-tuned from Latent Diffusion, or trained from scratch?
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.
Confidence
4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
Soundness
2 fair
Presentation
3 good
Contribution
2 fair
Limitations
The author has briefly addressed limitation. As a minor note, I recommend authors to include some qualitative results for limitation or challenging cases to encourage future research. I strongly urge the authors to consider two key improvements during the revision process: (1) If the authors intend to concentrate on image editing, it would be advisable to narrow down their focus to a single specific task. This would enable a more comprehensive exploration of the chosen task and provide greater clarity to the readers. (2) If the authors aim to emphasize the modality aspect, I recommend conducting a more extensive comparison with existing works. This would help establish the unique contributions of the proposed method and demonstrate its superiority over previous approaches. Regardless of the chosen direction, I believe that implementing these revisions would significantly enhance the paper's quality. --- **After Rebuttal**: I have read authors' rebuttal and other reviews. I thank authors for efforts to further address my concerns. After careful consideration, I decided to maintain my rating as 5-Borderline accept. My justification for not higher rating are: 1. The contribution about User's Interface (hybrid context injection strategy) is not (yet) fully validated. 1. Overlapping claims with Visual Prompting [1] paper (Also raised by [Reviewer zhtQ](https://openreview.net/forum?id=EmOIP3t9nk¬eId=d3daWKNPI3)) 1. Some Figure, Table need clearer explanation or captions (Also raised by [Reviewer RQEf](https://openreview.net/forum?id=EmOIP3t9nk¬eId=SRf3GZGiV2))
Summary
The authors propose a novel, visual-based method for image editing, using a pair of example images as instructions for users to specify the desired operations. The proposed idea uses the details and nuance in the example pairs to accomplish complex editing goals. The authors formulate this approach as a diffusion inpainting task, where the instruction pair and the query image are given and tiled as a grid. To achieve the desired quality, the visual prompt and region of user interest are injected into the diffusion structure as conditions, and the whole pipeline is trained as a latent diffusion model with classifier-free guidance. The authors also developed a metric to assess the quality of manipulations performed and conducted quantitative and qualitative studies.
Strengths
1. The proposed method demonstrates a new way to manipulate images using visual instructions only. 2. The model generalizes well on downstream tasks compared to other methods requiring task-specif models. 3. The authors propose a new metric to evaluate image manipulation tasks.
Weaknesses
1. For many common image editing tasks, it will be challenging or tedious for users to find the example pairs. For example, if a user wants to convert a photo into a painting by a specific artist, preparing the paired picture for the given painting will be hard without using existing pre-trained style-transfer models. 2. The authors think one of the disadvantages of using text prompts is that text prompts are abstract. However, the proposed method is also abstract if users want specific, exact manipulations. For example, if we're going to make the spout of a teapot 2 cm longer or want a particular proportion of the teapot shape, finding the pair examples and ensuring the model does the exact manipulations will be challenging. 3. Conducting user studies to support the assumptions would be great. For example, is it challenging for the users to find the pairs for their desired editing? Do the results look better to them (FID and CLIP scores are shown not to reflect human preferences/aesthetics in prior works like [1])? [1] Better Aligning Text-to-Image Models with Human Preference by Wu et al.
Questions
1. It is unclear how the image grid, Xt, is represented. Figure 2 shows that only I' is converted to noise after the forward diffusion process. However, in LDMs, the diffusion process is in the z space, where the whole image is encoded. 2. Is resolution one of the limitations of this method since the inpainting area is only 1/4 of the input image grid? 3. How the model is trained using the four datasets mentioned in 4.1 is unclear. Is it trained from scratch? 4. I am curious about how necessary the inpainting setup is. For example, what will the model perform if the task is to generate only I' instead of the whole grid (in other words, the visual clues and instructions are comping from cross-attention only)? 5. Have the authors tried incorporating the region of interest using different methods (such as masking in latent space similar to eDiff-I)?
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.
Confidence
2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
Soundness
3 good
Presentation
3 good
Contribution
3 good
Limitations
Yes, the limitations are addressed.
Summary
This paper introduces a novel image manipulation framework called ImageBrush, which allows for accurate and efficient image manipulation without the need for complex language descriptions. The framework consists of a diffusion-based generative model coupled with a hybrid context injection strategy, which facilitates better correlation reasoning. The authors also introduce a bounding box incorporation module to take user interactivity into consideration. The framework is evaluated on various downstream tasks and exhibits robust generalization abilities. Overall, the results demonstrate that ImageBrush generates compelling manipulation results aligned with human intent and paves the way for future vision foundation models.
Strengths
- The paper introduces a new image manipulation framework called ImageBrush, that does not rely on complex language descriptions. This may help mitigate severe prompt engineering efforts in applications. - A diffusion-based generative model coupled with a hybrid context injection strategy is proposed, which is a technically rigorous approach to image manipulation. - Extensive experiments are conducted on various downstream tasks and discuss this work's generalization abilities. - The paper is well-written and easy to follow, with clear explanations of the technical concepts and experimental results.
Weaknesses
- Although this work utilizes an in-context learning strategy to avoid heavy prompt engineering effort, the image prompt itself seems to arise another challenge. Take Fig. 4 as an example, the provided instruction (tiger pairs) itself necessitates heavy editing skills, that require another expert (either human or modern algorithms) to provide a snowy tiger. I'm worrying that this will not be practical in real scenarios given users mostly can not provide such a good pair of instructions, and writing prompts would be relatively easier. A more practical way is to provide a tiger with another snowy example (not have to be a tiger) so that it will be easier to collect inputs. However, this paper does not show this demo. - Besides the input collecting issue, it seems that the instruction and the input image are very similar in all provided results. If that is the requirement of making the proposed algorithm succeed, it will be further difficult to be applied to real scenarios, since one can not pre-collect corresponding instructions due to the in-the-wild testing case is not predictable. - The authors propose a new evaluation metric in Sec 4.4. I agree that instruction-based editing needs a reliable quantitative metric to help justify the performance of each algorithm, but I doubt whether it is a good idea to put it as one of the contributions. Proposing a new metric normally requires a large amount of user study that can faithfully provide evidence to show that the new metric is aligned with human perception, which is lacking in this work. I understand that will cost a huge effort and would probably produce a completely new submission. Therefore, I would suggest the author not claim this as a "novel" finding and lower its importance in the manuscript.
Questions
N/A
Rating
4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.
Confidence
5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
Soundness
3 good
Presentation
4 excellent
Contribution
2 fair
Limitations
See weaknesses.
Summary
In this paper, the authors propose a new method that can perform visual prompting via a pair of exemplar images. The authors introduce and train a new model for this task and evaluate the model performance on various datasets, showing strong empirical performance and a wide range of applications. They also propose new evaluation metrics to evaluate their method.
Strengths
This model introduced in this paper solves a very interesting and important application problem that is exemplar-based visual prompting of image generative models. The results look very promising and this model can be easily directly applied in a large range of intuitive applications by daily users.
Weaknesses
1. This paper is very similar to the paper “Visual Prompting via Image Inpainting” (https://arxiv.org/pdf/2209.00647.pdf). However, the authors didn’t compare their method with this paper. 2. The quantitative evaluations are not very comprehensive. For example, for tasks like semantic segmentation masks to image, the authors could have reported some standardized metrics like IoU. 3. Similar to (2), quantitative analysis on user-specified regional editing can also be included by comparing visual distances such as LPIPS.
Questions
1. How general is ImageBrush? How does the model perform on more OOD tasks/inputs (i.e. the prompts to which similar examples are not seen in training)? Or does it require training for each task? 2. How similar does the visual prompt and query image should look? Can the authors give some additional ablation study on the similarity between the visual prompt and thequery images? 3. Related to Weakness (1), can the authors compare their method with Visual Prompting via Image Inpainting? 4. Related to Weakness (2) and (3), can the authors report some standardized metrics for tasks like semantic segmentation mask to image, keypoint to image and user-specified regional editing? ~~I am happy to raise my score if the authors add experiments to address all my concerns.~~ ***After the rebuttal discussion, I would like to raise my score from 5o to 7.***
Rating
7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.
Confidence
5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
Soundness
3 good
Presentation
3 good
Contribution
3 good
Limitations
Although presented in the paper, the discussion of limitations and potential ethical issues is very limited. I would encourage the authors to extend this discussion by providing some failure cases and suggesting some future plans for safeguards, etc.
Summary
The paper proposes a novel image manipulation technique based on diffusion and visual context images. The authors argue that context from language can only communicate abstract ideas while giving visual context can bring across inexpressible concepts like art style or complex appearances. The novel technique adopts the 4-image-quadrant concept where the 4th quadrant is in-painted by to model to match some target (similar to SegGPT and Painter). But instead of in-painting segmentation masks, depth maps, or key-points in the 4th quadrant, they in-paint an edited version of the query based on transformation style or content given the context (prompts). To improve the grounding process, this work proposes to add an additional prompt encoder branch of the architecture which is fed into the diffusion process via Cross-Attention. This branch also allows optional guidance through bounding boxes on which the in-painting process should add additional focus. Bounding boxes can either be provided by drawing them or via language by using Grounding Dino. Finally, this work introduces a novel evaluation metric for exemplar-based image manipulation which does not need ground truth or human evaluation.
Strengths
- The paper is well-written and mostly easy to read. - The proposed approach is novel and the paper does a good job of highlighting the motivation for this line of work. The method is evaluated on multiple benchmarks and shows promising results. Comparisons are made with other image manipulation methods wherever possible. - The work is visually attractive with many examples of generated images and comparisons with similar methods - The authors highlight limitations of their approach by showcasing examples where the model works well but also cases where the model breaks down and does not work well. - Authors show qualitative and quantitative results on multiple datasets, showcasing the superiority of their method.
Weaknesses
- The description of the design of the prompt encoder is minimal and it would be interesting to have a more in-depth description. Specifically to how the different context image embeddings are combined. Are they just concatenated like the bounding box embeddings? - The section on the novel evaluation metric is somewhat unclear, specifically why the two directions of evaluation/combination of query and prompts are chosen. (Maybe more descriptive variable naming could help) - Generally, captions of figures and tables are minimal and contain little information. Specifically, figure 6 is confusing and could use an explanatory caption. - Equations 4 to 6 are unclear and not understandable. - Supplementary material is not formatted as an appendix.
Questions
Would the architecture still work if you would drop the 4-quadrant concept for two quadrants and only injecting context/prompts through cross attention?
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.
Confidence
4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
Soundness
3 good
Presentation
2 fair
Contribution
3 good
Limitations
Authors adequately address limitations.
I have read your rebuttal. Thank you for your efforts in conducting additional comparisons to Visual Prompting [1]. However, I still have several concerns. **(1) Unfair comparison to Visual Prompting [1]** This question is also raised by Reviewer [#zhtQ](https://openreview.net/forum?id=EmOIP3t9nk¬eId=d3daWKNPI3). The authors have added additional quantitative and qualitative comparisons to Visual Prompting (as in the Rebuttal PDF file). However, I am doubtful about an unfair comparison. As far as I understand (maybe the authors can also double-check), *Visual Prompting [1] was not trained on any "in-the-wild" datasets* (e.g., ScanNet [2], UBC-Fashion [3], and Image Editing [4] datasets). Would it be unfair to Visual Prompting, as it was not trained for the Image Editing task? (Figure 4, Rebuttal file). Moreover, I was surprised that Visual Prompting [1] (without being trained on any new tasks) performed reasonably well on most tasks compared to ImageBrush (e.g., on the Scannet dataset - PSNR: ImageBrush: 30.38 > Visual Prompting 29.34; on Inpainting - LPIPS: ImageBrush 0.816 > Visual Prompting: 0.787) (Table 1, Rebuttal file). These numbers are so close that it raises my concern. **If we train Visual Prompting [1] on the same dataset that was used to train ImageBrush, would Visual Prompting achieve a similar level of performance, or possibly even surpass ImageBrush?** (If the answer is yes... *Then what is the advantage of ImageBrush except from training on a larger datasets?* It is also worth to note that ImageBrush was finetuned from Latent Diffusion, which was trained on billions examples of [LAION dataset](https://laion.ai/blog/laion-5b/)) **(2) Figure 3 is still unclear to me**. I don't understand the quantity of the red and green columns. Is that the number of images that match the specific threshold on the x-axis or y-axis? How many images are used to plot this Figure? Where do these images come from? Are these images used in training? **(3) Doubt about Interface design.** I am still left with questions about Figure 1 in the Rebuttal PDF file. 3.1. What's the green-like bounding box in every example image? 3.2. The authors still haven't answered my question: "I feel like these results can be generated by different noises. For example, given the same visual prompt (e.g., "Put a crown on the woman"), **different noises can also lead to different visually different crowns.**" Is this statement true? To be specific, using the same example-pair and query image, is it true that I can generate different outputs by applying different noises? In other words, can different noises lead to different images, each focusing on a different aspect (e.g., one image focusing on the hat, other images focusing on the beard)? Please refer to [Figure 12 in the InstructPix2Pix paper](https://arxiv.org/abs/2211.09800) for a visual example of what I am trying to depict. My other unanswered questions are as follows: **(4) Unfair comparison in video inpainting task**. "Following [4], Video-FID [...] is used to evaluate the methods. Would it be better to use Video-FID instead of FID?" **(5) Questions about the focusing goal of this paper (Stated in Limitations)**. I want to know how the authors identify this work. Will it be a novel method for image editing, or will it be a powerful modality framework (e.g., capable of solving a wide range of tasks)? Thank you! *Reference:* [1] Amir et al., *Visual Prompting via Image Inpainting*, NeurIPS 2022. [2] Angela et al., *ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes*, CVPR 2017 [3] Polina et al., DwNet: *Dense warp-based network for pose-guided human video generation*, BMVC 2019 [4] Tim et al., *InstructPix2Pix: Learning to Follow Image Editing Instructions*, CVPR 2023.
**_Unfair comparison to Visual Prompting [1]._** 1) For fair comparison we have carefully re-trained [1] using the same dataset we employed for ImageBrush. Without training on the same dataset, their performance is super terrible, which is uninformative and thus not shown. 2) Actually, both Table 1 and Figure 4 illustrate their performance after training on the same tasks as ours. I guess that’s why it feels reasonably well. From Table 1, we could see that even training on larger datasets as ours, their results are still below ImageBrush. Additionally, Figure 4 reveals that they tend to produce conservative results and lack details. Although it is a good strategy to attain decent metric value, their results are not photorealistic, which will largely limit the real-world application. **_Clarification of Fig. 3_** 1. The quantity of green and red columns does not correspond to the count of images. Its height reflects the average value (CLIP Direction/Image Similarity) derived from all the images encompassed within the respective bins. 2. As for the bins, consider the left figure as an illustration. For the first green and red columns sitting around 0.8, we compute the mean CLIP Direction Similarity for all images falling within the CLIP Image Similarity range of $[0.8-1/6, 0.8+1/6]$. 3. To draw this figure, we employ a dataset of 10k images that remained unseen throughout the training phase, as stated in the Dataset of Section 4.1. **_Doubt about interface design._** 1. We apologize for the oversight in visualization. It appears that we inadvertently depict the bounding boxes produced by Ground DINO. 2. After a close examination of Figure 12, we concur that diverse noise sources can indeed yield varied outcomes due to the intrinsic nature of DDPM. But utilization of bounding boxes could further steer the model's attention towards the highlighted area. This effect is naturally facilitated by the statistical correlation between bounding boxes and relevant editing regions. To elaborate, Ground DINO enables the automatic identification of the most pertinent spatial region within the image. This explicit indication of the editing intention furnishes our model with an interface to more effectively capture such intentions. Our statistical observations and qualitative examples of Rebuttal file validate its proficiency in capturing intention. 3. Regarding the interplay between bbox and noise effects, we recognize that this deserves further exploration. In future, we will show more examples with less ambiguity to demonstrate its efficacy. We really appreciate your constructive insight. **_Unfair comparison in video inpainting task._** 1. For video inpainting tasks, the Video-FID stands as a crucial metric to evaluate temporal aspects of video synthesis. Here we include a comparison with the strong baseline TSAM(ImageBrush 175.25 vs. TSAM 158.30). ImageBrush achieves close performance even without special designs to harness temporal information. 2. Notably, the computation of the Video-FID relies on the i3d feature extractor, which has been pre-trained on an extensive collection of videos. Consequently, the use of smoothed video features is favored. But for our approach, which lacks temporal awareness, the extracted features are likely to diverge significantly from their distribution. As for the FID metric that evaluates each frame independently, the features extracted by the Inception model tend to align closely with the Ground Truth. This discrepancy in feature behavior explains why our model attains a higher FID score while yielding a comparatively lower Video-FID score. **_Questions about the focusing goal._** 1. Characterizing this work as a novel image editing approach may not be entirely accurate. The selection of image editing is rooted in the fact that image manipulation contains abundant contexts, offering potential solutions for practical real-world applications. 2. We prefer to locate our work as a visual in-context learning investigation from the perspective of image generation, which is complementary to VisualPrompting. While they show potential in visual in-context learning, their focus remains on recognition like detection and segmentation, thereby having limited applications. But directly extending their method to generative tasks will encounter challenges on capturing editing intention and produces conservative synthesis, as Fig.4. Therefore, we introduce the diffusion mechanism, which guarantees intention comprehension and realistic synthesis. Our work incorporates a meticulously designed prompt encoder to grasp the underlying manipulation intention and an interface to enhance its understanding. In this way, we open avenues for novel applications (see Common Response). This extension significantly expands the range of potential applications and signifies a substantial step towards establishing a foundational model for vision-based applications.
Thank you for providing clarification. All my concerns about unfair comparisons have been resolved. I understand Figure 3 now, but I'd like to offer a minor note. Since I didn't understand the concept of Figure 3 without your explanation, it might be possible that readers find it hard to understand too. My suggestion would be a histogram-like graph (e.g., [Figure 7 in Visual Instruction Inversion paper](https://arxiv.org/pdf/2307.14331.pdf)), or a line graph (e.g., [Figure 8 of InstructPix2Pix](https://arxiv.org/pdf/2211.09800.pdf)), for a clearer interpretation. However, the choice of the final plot style is up to authors. It's unfortunate that the Interface module requires further thorough investigation. I second the potential of this module, as it could open the way for a novel interactive system for image editing. Finally, I (personally) share the same view as the authors regarding the focus of the goal of this work: visual in-context learning. --- I will read other review again and finalize my rating. Thank you!
Thank you for your constructive suggestions, which have been very helpful to us. We appreciate your understanding and recognition. Indeed, we have realized the confusion in Figure 3, and we plan to reconstruct it by referring to Figure 7 in the [Visual Instruction Inversion](https://arxiv.org/pdf/2307.14331.pdf) [1] paper, making it easier to understand. We will also revise the description of our motivation to have a clearer focus and include more validation results for the Interface module in subsequent versions. As for the novel interactive system for image editing, we plan to open-source our demo and code in the future, allowing more users to participate and try it out. This is crucial for iterating and exploring a more user-friendly image editing application paradigm. In regards to visual in-context learning, we hope that our work can inspire more follow-up studies to explore new ways of using visual generative models and more general application scenarios. *Reference:* [1] Nguyen et al., *Visual Instruction Inversion: Image Editing via Visual Prompting.* arXiv preprint arXiv:2307.14331.
Thank you for the response!
Thank you for the response! You have addressed my concerns/question and given the rebuttals to other the reviewers, I will raise my score.
Thank the authors for the rebuttal. Based on our discussion above, I would suggest the authors to clarify the limitation and use cases of the proposed method in the revision. I will raise my score to 6.
Decision
Accept (poster)