We would like to thank you for the positive feedback, helpful comments, and the support of our work. Following are our responses to each individual comment (which are highlighted in italics).

Response for Weaknesses (RfW)

W1: I think the discussion of limitation is too shallow. I recommend providing more failing cases and discussing the potential directions to enhance the results.

RfW1: Thank you for the constructive suggestion. We have provided more failing cases to illustrate the limitations of our study, as shown in Fig. 1 and Fig. 2 of the attached PDF with our author rebuttal. We also provided in-depth discussion about these limitations for potential future directions, please see the detailed analysis of the ''Limitation'' section in General Response to reviewers.

Response for Questions

Q1: I wonder why the interpolation is called ''Bilateral'' nearest neighbor interpolation. For me, ''Bilateral'' term should be used for a scheme that considers the value as well. And the proposed interpolation scheme is just find the nearest null value. This is not critical but I think it is misleading.

We apologize for any confusion caused by our terminology. The term ''Bilateral'' was chosen to emphasize that the interpolation considers nearest neighbors in two directions (i.e., both the X and Y axes), highlighting that the interpolation process involves adjacent non-null data across two dimensions. We will explain this terminology in the camera-ready version for better clarity.

Q2: It is unclear to me what the brighter region around the drag arrow means. Is that a mask? Are those region/mask included and used in drag instructions?

Yes, you are correct, the brighter region is a mask, indicating the image area to be edited. It is used as drag instruction with the drag arrows, which is widely adopted in drag-based editing [4, 18, 19, 27].

We will explain this in the camera-ready version for better clarity.

Q3: I am confused about Fig.8. What is the desired effect and result of the drag edit? Semantically, the user is moving ''the dog’s'' hand, not ''the cat’s''. However, the result seem to generate cat’s hand for all different steps. Is this some limitation of all drag-based method? If yes, I think it is worth discussing.

We apologize for any confusion regarding Fig.8. This case was randomly selected to illustrate an ablation study on number of inversion steps in terms of drag effect, which is used to determine the number of inversion steps in diffusion inversion. It is not intended to demonstrate a specific editing result. To avoid such confusion, we will replace Fig. 8 with a more proper case in the camera-ready version.

Regarding your question, the observed phenomenon is common for diffusion models. In this case, when ''the hand'' is moved onto ''the cat'', it semantically associates with ''the cat'', resulting in the generation of ''the cat’s hand''. This is because the U-Net architecture used in diffusion models contains numerous CNN layers, which fuse features of the cat and the hand in the latent space when their positions are close together. For generative artificial intelligence, the ability to generate image with contextual semantic association is actually considered a strength of diffusion models, such as Stable Diffusion.

Q4: I wonder if there are metrics or evaluation processes that can evaluate the content preserving quantitatively. Regarding image editing, I think this is a very important factor and the current paper only provides a couple of images as evidence.

Indeed, we have provided evaluation results with regard to content preserving in terms of ''1-LPIPS'' metric, as shown in Table 1 of our paper. ''1-LPIPS'' metric is widely used to for image content consistency evaluation for drag editing [4, 14, 15, 27]. Our experiments are conducted on DragBench dataset, and the results can demonstrate the effectiveness of our method, as explained in general response (i.e., evaluation metrics and statistical rigor) to reviewers.

Response for Limitations

As I mentioned above, I think the limitation discussion is not enough (I also raised another potential limitation). Moreover, I strongly recommend including the limitation discussion in the main paper, not supplemental material.

Thank you for the constructive suggestions. Due to page limitation, we did not include limitations in the main paper of the submitted paper. However, we acknowledge the importance of this discussion and will try our best to enhance and include it in the main paper of the camera-ready version.

We would like to thank you for the positive feedback, helpful comments, and high praise and recognition of our work. Below are our responses to each individual comment (highlighted in italics).

Response for Weaknesses

W1: Theoretical Depth

Our one-step optimization strategy is developed by simulating strain patterns in stretched materials, inspired by [20]. The core of this strategy is LWF, which is strictly derived by formula (3)-(10) in our paper.

Furthermore, our effort in this study is to significantly reduce the editing time, the experiments we conducted have fully verified the contribution we claim. Specially, the results in Table 1 of our paper show FastDrag's ultra-fast editing speed compared to SOTA methods.

Additionally, the principle of BNNI is based on the most fundamental interpolation principles in the field of image processing.

W2, 3 and 4: Experimental Design, Statistical Rigor and Comparison with State-of-the-Art

The evaluation methods on DragBench [27] dataset are widely used in the field of drag-based editing, following the state-of-the-art methods [4, 14, 15, 27]. The results presented are averages across DragBench, which includes diverse image categories and drag tasks. Thus, the results on it are sufficiently demonstrating FastDrag’s superiority and robustness.

Additionally, for Statistical Rigor, please refer to the ''Statistical Rigor'' of the General Response.

Regarding the trade-offs between speed and quality, please see the response RfW3 to Reviewer eL9M.

W5: User Study

Thank you for your valuable suggestions. We plan to conduct a user study through surveys in the released project demo to ensure the reliability of the results. The outcomes of this study will be published following the demo release.

W6: Limitations Discussion

Regarding Limitations, please refer to ''Limitation'' in General Response to reviewers.

W7: Detailed Code Instructions

Thanks for your valuable suggestion. We will improve it and provide more detailed tutorials in the released code.

W8: Ethical Considerations

It is a common concern with generative methods. We will enforce strict open-source licensing in the publicly released code to limit unethical use.

W9: Technical Details of U-Net and LCM

Our diffusion structure's implementation entirely follows the mainstream baseline DragDiffusion [27]. Specifically, the U-Net used in our model is widely used in image generation [4, 14, 15, 27, 31]. As stated in Appendix B, the U-Net structure is adapted with LCM-distilled weights from LDM (i.e., Stable Diffusion 1.5). We will further clarify this in the camera-ready version.

W10: Long-Term Performance

Thank you for your interesting idea. We conducted experiments and found that repeated editing using diffusion models can degrade image quality over time due to accumulated errors. This may be mitigated by refined training techniques or corrective algorithms. Due to PDF space constraints, we are unable to include the figure in this rebuttal but will include related discussions in the camera-ready version.

Response for Questions

Q1: Latent Warpage Function (LWF) Details

In Section 3.2.1, we provide a detailed explanation of the derivation process for the Latent Warpage Function (LWF).

1. We model the stretching behavior by Equation (3), which normalizes and aggregates ''component warpage vectors'' $p_jp_j^{i\ast}$ caused by multiple drag instructions into a single warpage vector for subsequent latent optimization, as shown in Fig. 3. The $p_jp_j^{i\ast}$ are modeled by Equation (5), with the stretch factor modeled by Equation (6), designed to simulate the behavior of stretched materials.

2. The degree of pixel adjustments is inversely proportional to the distance of the pixel from handle point $s_i$.

3. The direction of pixel adjustments is the sum of $p_jp_j^{i\ast}$, whose direction is the same as corresponding drag instruction's direction $|\overrightarrow{s_i e_i}|$.

Q2: BNNI in Complex Scenes

Thank you for your valuable opinion. BNNI may struggle with cases involving detailed textures, as discussed in our response RfW1 to Reviewer eL9M. However, BNNI effectively handles most cases. Results on the DragBench dataset demonstrate its comparability with SOTA methods. We will try to improve it in future work.

Q3: Consistency-Preserving Strategy Details

Key-value pairs contain semantic information through attention mechanism during diffusion inversion, which has been proved by many researches [2, 18, 19]. Via cross-attention mechanism in each step, key-value pairs containing ori-image's semantics guide the sampling process to maintain consistency. We use the ''1-LPIPS'' metric to evaluate consistency preservation. Please refer to ''Evaluation Metrics'' in General Response.

Q4: Scalability with Image Complexity and Dataset Size

There is no existing drag-based image editing benchmark with a focus on intricate images or larger datasets. Although DragBench is not as large as some traditional datasets, it is the most complex dataset in its field—featuring more classes and a mix of real and generated images compared to those used in [14, 21, 27], which contain less image number and class or miss key drag instructions, such as mask. It also supports our primary contribution of improving editing speeds.

Q5: Negative Societal Implications

Thanks for your valuable suggestion. About this issue, please see response in RfW8.

Q6: Limitations and Future Work

Regarding Limitations and Future Work, please refer to ''Limitation'' in General Response to reviewers.

Q7: Sensitivity to Initialization

FastDrag uses DDIM inversion strategy to get the initial state of the latent space as same as other drag-based editing methods [14, 15, 31]. Since our latent optimization is decoupled from the diffusion inversion and sampling process, Fastdrag is not sensitive to the initial state of the latent space.

We would like to thank you for the positive feedback and helpful comments. This rebuttal addresses your comments and suggestions for conciseness. Following are our responses to each individual comment (which are highlighted in italics).

Response for Weaknesses (RfW)

W1: ... needs to include more comprehensive evaluation metrics ...

RfW1: Regarding evaluation metrics, following the SOTA methods [14, 15, 27], we employ ''MD'' (Mean Distance [22]), ''IF'' (i.e.,''1-LPIPS'' [10]) as evaluation metrics for a fair comparison. The metrics suggested by Reviewer, such as SIFT, are primarily used for feature point detection and matching. The suggested metrics may not be appropriate for drag-based editing due to semantic changes within the editing region. It is because that drag editing requires measuring the similarity between the edited result and the desired semantics, rather than comparing the ground truth and the edited result.

Due to the space limitation, please refer to ''Evaluation Metrics'' in the general response to reviewers, where we conduct more detailed analysis and explanation.

W2: Miss standard error and it is unclear if the result are ...

RfW2: Our method is evaluated on the widely used DragBench dataset [27] for drag-based editing, following the state-of-the-art methods [4, 14, 15, 27] for a fair comparison. The results presented are the averages across DragBench that includes diverse image categories and drag tasks, which are not cherry-picked or randomly chosen. This has been explained in the General Response to Reviewers, please refer to the Evaluation Metrics in general response to reviewers.

Regarding standard errors, to address your concern, we conducted an additional experiment by repeating our experiment 10 times under the same experimental settings. We observed that the variances of the performance metrics obtained from 10 realizations of our FastDrag are MD (0.000404), 1-LPIPS (9.44E-11), and Time (0.018), all of which fall within a reasonable range.

W3: ... effect of the edits on "other" parts of the image ...

RfW3: There are almost no changes outside the masked editing region using our method. It is because our approach optimizes the latent strictly within the masked region, as described by Equations (8)-(9) in our paper. We also employ a consistency-preserving strategy to maintain the consistency of image content outside the mask region. Note that, this is almost the same for other drag editing methods, as they also conduct latent optimization within the masked region, and adopt different strategy to maintain consistency such as LoRA [14, 15, 27, 31].

W4: ... smooth and lose finer details ...

RfW4: The limitation of smooth and losing finer details is common across diffusion-based drag methods, inherent to the diffusion models employed, such as LDM and LCM. This issue arises due to model approximations, randomness, and potential computational errors, which can result in imperfect symmetry between the inverse and sampling processes, leading to a situation where generated image may not be exactly the same as the original [28].

For instance, DragNoise [15] and DragDiffusion [27] introduce textures not present in the original images, affecting fidelity, as illustrated in Fig. 6 of paper. However, although our method may result in finer details loss, it outperforms others in overall task execution, especially in editing speed. We will discuss the limitation in the camera-ready version. Future research will be conducted to mitigate such effect and enhance the editing performance.

Response for Questions

Q1: Does the reported computational time include the inversion process?

Yes, the reported computational times include the inversion process, sampling process, and the time for latent optimization.

Q2: What happens if you integrate DragDiffusion with a single-step diffusion model? How does it compare with Fastdrag?

When integrating DragDiffusion with a single-step diffusion model, the editing time is still much longer than using FastDrag. For DragDiffusion and FastDrag under diffusion steps of 1, 20, and 50, we calculate the time required for inverse, sampling, and latent optimization respectively. The results provided in Fig. 3 of the attached PDF show that even with a single diffusion step (i.e., diffusion step set as 1), DragDiffusion still requires significantly more time (20.7 seconds) compared to FastDrag (2.88 seconds).

Q3: Why does iterative drag diffusion suffer compared to the one-step approach?

These methods require $n$-step iterations to achieve semantic optimization, with each step to optimize semantics within small editing area of the image, thus they need $n$ small-scale, short-distance optimizations to achieve overall latent optimization, and require large amount of time to perform $n$ iterations of the optimization. Whereas our method only requires a single short-time computation on the latent to achieve the semantic optimization, thereby significantly reducing editing time.

Q4: How does FastDrag affect other parts of the image during edits?

During the editing process, other parts of the image outside the mask remain almost unchanged, as explained in RfW3 to the reviewer.

Q5: Can you perform multi-point drag editing with FastDrag?

Yes, our method can perform multi-point drag editing, which has been demonstrated in Fig. 6 (row 4) of our paper and Fig. 13 (row 6) in Appendix C.

We will provide further clarifications and explanations regarding RfW2 and Q3 in the camera-ready version of our paper.

We would like to thank the reviewer for the positive feedback and helpful comments. Following are our responses to each individual comment (which are highlighted in italics).

Responses for Weaknesses (RfW):

W1: If the drag distance is long, will the BNNI still success to maintain high semantic quality, how about the editing speed and complexity for long-distance dragging and latent relocation?

RfW1: For typical drag tasks, dragging is usually performed over short distances. However, for long-distance dragging, which is generally an object moving task, as illustrated in Figure 13 (row 3) in the Appendix, we do not need to employ the BNNI strategy. Instead, we can fill the semantics around the target location or a manually designated area of the image into the moved object's original location [18, 19], as described in Appendix B of our paper. We will further clarify this in the camera-ready version.

However, for extremely long-distance drag editing, our method may lose some details of the object, as explained and discussed in the "Limitation" of our general response to reviewers. Please refer to the limitation regarding "Extremely Long-distance Drag Editing" in our general response to reviewers.

Nonetheless, long-distance editing does not increase the editing speed and complexity of our method. Our method can achieve latent optimization with one-step warpage optimization using Equations (3)-(10), which is not affected by the editing distance.

W2: There should also be some failed examples to better illustrate the proposed method.

RfW2: Thank you for your valuable suggestion. We have included failed examples to illustrate the limitations of our method for better understanding, as shown in Figures 1 and 2 of the attached PDF with our author rebuttal. Please refer to the limitation discussion in our general response to reviewers.

W3: I am curious whether using a better base model can achieve better editing results, and whether there is a trade-off between editing time and editing performance.

RfW3: Regarding better base model, it may potentially enhance editing results. Since the compared state-of-the-art methods all utilize SD 1.5 for drag tasks, our experimental results are solely based on SD 1.5. Theoretically, our one-step optimization method is independent of the base model. Therefore, FastDrag can perform drag tasks based on more advanced models such as SDXL or SDXL turbo, and its editing performance theoretically varies with the base model used. However, integrating more advanced models like SDXL would require substantial code base restructuring (we use relatively early version of diffusers), due to time constraints in the rebuttal period, we are not able to implement this in time. We will support more base models in the future.

Regarding the trade-off between editing time and performance, there should be a trade-off between editing time and performance for $n$-step optimization methods, as they all require $n$-step iterative optimization to achieve desired editing performance, though these methods can reduce editing time by decreasing the number of iterations, this typically degrades performance, since less iteration will lead to insufficient latent optimization. Thus there should be a trade-off for $n$-step methods.

In contrast, FastDrag leverages one-step warpage optimization to achieve latent optimization via Equations (3)-(10), the editing time will not be influenced by editing task or images. Therefore, our method does not have the trade-off between editing time and performance.

W4: More recent works [ref1] should be included for comparison.

RfW4: Thank you for the suggestion. The study [ref1] did not publish or preprint when we submitted FastDrag, thus we did not compare it. We will discuss and compare it in our camera-ready version.

[ref1] EasyDrag: Efficient Point-based Manipulation on Diffusion Models, CVPR 2024