GoodDrag: Towards Good Practices for Drag Editing with Diffusion Models

We thank the reviewer for the valuable comments on this paper and address the raised issues here and in the revised manuscript.

Q1: DragonDiffusion and DiffEditor

A1: Thank you for highlighting the related works, DragonDiffusion and DiffEditor. While these methods involve image editing within the denoising diffusion process, they are fundamentally different from GoodDrag and the proposed AlDD in both theoretical foundations and practical implementation.

From a theoretical perspective, DragonDiffusion and DiffEditor rely on a mechanism analogous to classifier guidance. In these methods, the diffusion process remains probabilistically grounded, and each step is guided by combining the unconditional gradient and the conditional likelihood term. Mathematically, this is expressed as:

$\nabla_{x_t} \log q(x_t \mid y) = \nabla_{x_t} \log q(x_t) + \nabla_{x_t} \log q(y \mid x_t)$

where the guidance operates within the probabilistic framework of diffusion (see Eq. 8 of DragonDiffusion and Eq. 3 of DiffEditor). From a practical perspective, this results in the editing and denoising processes being intertwined and inseparable.

In contrast, GoodDrag follows a fundamentally different paradigm, similar to DragDiffusion, where the drag editing operations and the denoising diffusion process are decoupled. AlDD distributes drag operations strategically across multiple diffusion steps but is not constrained by the probabilistic formulation of classifier guidance, which represents a significant departure from existing methods. This separation allows AlDD to introduce flexibility in drag editing, which is not feasible with methods like DragonDiffusion and DiffEditor, and effectively improves the results. As shown in Section 5.2 and Appendix B-C, GoodDrag achieves significantly better performance than DragonDiffusion across multiple benchmarks, both quantitatively and qualitatively. These results underline the practical advantages of AlDD and the distinctiveness of GoodDrag’s approach.

Q2: Runtime analysis.

A2: We include runtime analysis in Appendix L of our paper. Since the proposed Information-Preserving Motion Supervision (IP) involves $J$ motion supervision steps as introduced in Eq. 7, the runtime of GoodDrag is slightly longer than DragDiffusion (71.3s vs. 57.4s) as shown in the table below. For a better comparison, we modified DragDiffusion by increasing the number of drag operations to match the number of motion supervision steps used in GoodDrag. While this updated version (referred to as DragDiffusion*) requires a longer runtime, it still underperforms compared to GoodDrag as shown in the table below, highlighting the advantages of our approach.

Additionally, we tested a simplified version of our model without the IP component, relying solely on the proposed AlDD strategy. This variant (w/o IP) is significantly faster than DragDiffusion (32.1s vs. 57.4s) while still achieving better performance than DragDiffusion. These results further demonstrate the efficiency and efficacy of the proposed algorithm.

We thank the reviewer for the detailed follow-up and for raising this important concern. Here, we provide additional clarification on the differences between DiffEditor and GoodDrag, specifically addressing the novelty of AlDD.

Contrary to the claim that DiffEditor alternates between dragging and denoising during the diffusion process, the drag and denoising steps in DiffEditor are performed together, not separately. For further reference, please see Algorithm 1 on Page 5 of the DiffEditor paper. The editing operation in DiffEditor is executed via Line 17 in Algorithm 1, defined as:

$\mathbf{z} _ {t-1}=\mathbf{m} _ {edit}\cdot \mathcal{F}(\mathbf{z} _ t;\eta_1(t))+(1-\mathbf{m}_{edit})\cdot \mathcal{F}(\mathbf{z}_t;\eta_2(t))$

To more clearly see why drag and denoising are intertwined in DiffEditor, consider the expanded form of the first term $\mathcal{F}(\mathbf{z}_t;\eta_1(t))$ using Eq. 2 of DiffEditor (the second term $\mathcal{F}(\mathbf{z}_t;\eta_2(t))$ is similar and omitted for clarity):

$\mathcal{F}(\mathbf{z} _ t;\eta_1(t))=\sqrt{\alpha _ {t-1}} \frac{\mathbf{z} _ {t}-\sqrt{1-\alpha _ {t}} \tilde{{\epsilon}} _ {\theta}^{t}\left(\mathbf{z} _ {t}\right)}{\sqrt{\alpha _ {t}}} + \sqrt{1-\alpha _ {t-1}-\eta_1(t)^{2}} \cdot \tilde{{\epsilon}} _ {\theta}^{t}\left(\mathbf{z}_{t}\right) + \eta_1(t) {\epsilon}$

With Eq. 4 from DiffEditor, we further expand the formula as:

$\mathcal{F}(\mathbf{z} _ t;\eta_1(t))=\sqrt{\alpha _ {t-1}} \frac{\mathbf{z} _ {t}-\sqrt{1-\alpha _ {t}} \left[\epsilon _ {\theta}^{t}\left(\mathbf{z} _ {t}\right)+\eta\cdot\nabla _ {\mathbf{z} _ t} \mathcal{E}(\mathbf{z} _ t, \mathbf{y})\right]}{\sqrt{\alpha _ {t}}} + \sqrt{1-\alpha _ {t-1}-\eta _ 1(t)^{2}} \cdot \left[\epsilon _ {\theta}^{t}\left(\mathbf{z} _ {t}\right) +\eta\cdot\nabla _ {\mathbf{z} _ t} \mathcal{E}(\mathbf{z} _ t, \mathbf{y})\right] + \eta _ 1(t) \epsilon$

Here, ${\epsilon} _ {\theta}^{t}\left(\mathbf{z} _ {t}\right)$ represents the denoising signal, while $\nabla_{\mathbf{z}_t} \mathcal{E}(\mathbf{z}_t, \mathbf{y})$ represents the drag signal. These two signals are fused and inherently intertwined in DiffEditor’s editing process and cannot be easily decoupled.

This design is deeply rooted in the probabilistic framework of classifier guidance (Eq. 12 on Page 8 of "Diffusion Models Beat GANs on Image Synthesis" or Eq. 3 on Page 3 of DiffEditor), which mandates that the denoising and guidance signals be applied together.

It’s worth noting that the $t$%$2==0$ on Line 13 of DiffEditor’s Algorithm 1 might be mistaken for alternating drag and denoising. However, as explained on Page 6 of DiffEditor, it is to reduce the number of guidance timesteps, which we quote:

“Due to the guidance enhancement from regional guidance and time travel, we can achieve editing with fewer guidance time steps, i.e., we introduce gradient guidance every two time steps in sampling.”

This does not decouple the drag and denoising operations in $\mathcal{F}(\mathbf{z}_t;\eta_1(t))$.

In contrast, GoodDrag, through the AlDD framework, explicitly separates the dragging and denoising processes. This separation is not a superficial design choice but rather a foundational distinction that allows AlDD to strategically alternate between image editing (motion supervision and point tracking) and artifact reduction (denoising). This independence offers greater flexibility and precision in managing drag operations, which is not achievable in DiffEditor’s tightly coupled probabilistic framework.

Discussion on Related Work. We have cited both FastDrag and RegionDrag in the related work of our paper. While the deadline for updating the PDF has passed, we will include a more detailed discussion of these works in the final version.

We hope this clarification addresses the remaining concern, and we thank the reviewer for the opportunity to elaborate further.

We thank the reviewer for the valuable comments on this paper and address the raised issues here and in the revised manuscript.

Q1: Evaluation with IF and MD

A1: We provide the MD and IF results in Table 5 and 6 of the revised Appendix C, and also include them below for easier reference. Our method achieves better MD values than the baseline methods, demonstrating its effectiveness in drag editing. Note that while MD can well measure drag accuracy, it is considerably slower than the proposed DAI as detailed in Appendix C.

While the IF score of our method is slightly lower than other approaches, we argue that the IF metric is fundamentally flawed as an evaluation measure for drag editing approaches. IF is defined as 1-LPIPS between the drag-edited image and the input image, meaning it penalizes any changes to the image, even when such changes are necessary to achieve the desired editing. As a result, the metric rewards outputs that are identical or nearly identical to the input image, which contradicts the very purpose of drag editing.

This inherent limitation is clearly demonstrated in Figure 11 of the revised Appendix C, where an example with superior visual quality receives the worst IF score, underscoring its failure to reflect meaningful drag edits. In contrast, the proposed GScore metric better correlates with human perception, making it a more reliable and appropriate metric for evaluating drag editing algorithms. Please refer to Appendix C for more details.

Q2: Results on DragBench

A2: The results on DragBench are provided in Appendix B and C (Tables 3–6). Our GoodDrag demonstrates favorable performance compared to state-of-the-art methods, consistently achieving superior results across different metrics, including DAI, GScore, and MD.

Q3: Different base model

A3: The proposed GoodDrag framework is compatible with different diffusion base models. While we use Stable Diffusion 1.5 as the default model in this work, we also tested GoodDrag with Stable Diffusion 2.1 and observed minimal difference in performance, which demonstrates the robustness of GoodDrag across different base models. Please refer to Appendix K for more details.

Q4: Limitation

A4: Similar to existing diffusion-based methods, such as DragDiffusion, the proposed GoodDrag relies on DDIM inversion for effective drag editing. However, DDIM inversion may face challenges in complex scenarios, where the reconstruction of the inversed image appears blurred and many fine details are lost. Consequently, the edited result of GoodDrag also suffers from these artifacts. Please refer to Appendix N for more details.

Dear Reviewer k1h1,

As the discussion phase is coming to a close, we wanted to check if our responses have fully addressed your concerns. If there are any remaining issues that need clarification, please let us know—we’d be happy to provide additional details.

Thank you!

Dear Reviewer k1h1,

Thank you for your valuable and constructive feedback on our submission. Your comments have been instrumental in improving the quality of our manuscript.

As the discussion deadline approaches, we want to ensure that all your concerns have been fully addressed. Please feel free to reach out if you require further clarifications or additional analyses.

To address your points:

• Results for DragBench with MD and IF are in Appendices B and C, with comparisons for DAI, GScore, MD, and IF detailed in Tables 1–6.

• Different base models are discussed in Appendix K.

• A new section has been added in Appendix O to discuss the limitations of our work.

We hope these updates address your concerns and reflect the improvements made to the paper. If you have had a chance to review our rebuttal and revisions, we would greatly appreciate your feedback or confirmation on whether the updated manuscript meets your expectations.

Additionally, we hope you might consider raising your score based on the revisions, as we believe these updates, guided by your thoughtful review, have substantially improved the manuscript.

Thank you once again for your time and insights.

We thank the reviewer for the valuable comments on this paper and address the raised issues here and in the revised manuscript.

Q1: The novelty of the AlDD design is limited, as it mainly alters the order of denoising during the drag updating process.

A1: While the operation of AlDD may appear simple at first glance, its development required new insights and deep understanding of the underlying problem, and no prior work has proposed this approach. Notably, we for the first time demonstrate that a subtle adjustment in the denoising order can have a substantial impact on the drag editing process, which is by no means straightforward to identify or realize.

As evidenced in Figure 4, Figure 14, and Table 1, AlDD effectively improves the accuracy and quality of drag editing. Furthermore, GoodDrag, incorporating AlDD, achieves superior performance compared to SOTA methods across multiple datasets, including Drag100 and DragBench, which is demonstrated both qualitatively and quantitatively across various metrics, such as DAI, GScore, and MD, as shown in Figure 7–10, Figure 12–13, and Table 1–10. These results underscore the importance of AlDD, highlighting its innovative contribution to the field and its practical effectiveness in delivering state-of-the-art drag editing performance.

Q2: How to choose hyperparameter B

A2: As explained in our paper (Figure 3), $K$ represents the total number of drag operations, and $B$ denotes the number of drag operations performed before each denoising step during AlDD. Consequently, the total number of denoising steps in the AlDD phase is $K / B$, and the range of $B$ is constrained by $1 \leq K / B \leq T$ , where $T$ is the starting timestep, representing the total number of denoising steps in the drag editing process.

As introduced in Section 5.1, we set $T = 38$ and $K = 70$, which gives the range $2 \leq B \leq 70$. When $B = 70$, the process reverts to the baseline case without AlDD, where all drag operations are performed at a single timestep ($t = T$).

In general, $K / B$ should not be too large, as drag operations performed at very low timesteps ($t$ close to 0) occur in a latent space that closely resembles the image space, which is unreliable for significant edits. We experimented with various values of $B$ and found that $B = 10$ (resulting in $K / B = 7$ denoising steps in AlDD) consistently yields desirable drag editing results. Therefore, we use a fixed $B = 10$ throughout all our experiments.

Q3: “ALDD” and “AlDD”

A3: Thank you for pointing out the inconsistency in naming. “ALDD” was a typo error which we have corrected to “AlDD” in the revised manuscript.

Q4: Blank area on Line 054

A4: Thank you for pointing this out. We have revised the formatting to remove the blank area on Line 054 in the updated manuscript.

Q5: Diverse drag tasks compared to DragBench

A5: While DragBench contains more images, we argue that diversity cannot be solely measured by the number of images. As introduced in Section 4.1 of the main paper, our insight lies in recognizing the variety of drag editing tasks, which includes relocation, rotation, rescaling, content removal, and content creation, as illustrated in Figure 6. These tasks encompass distinct characteristics that DragBench does not explicitly address. Relocation involves moving an object or a part of an object, while rotation adjusts the orientation of objects; both tasks mimic rigid motion in the physical world without changing the object area or creating new contents. Rescaling corresponds to enlarging or shrinking an object. Content removal involves deletion of specific image components, such as closing a mouth, whereas content creation involves generating new content not present in the original image, such as opening a mouth. By thoughtfully incorporating samples of all different tasks, Drag100 prioritizes task diversity over sheer quantity, offering a more comprehensive and balanced evaluation framework for drag editing algorithms.

Q6: Results on DragBench

A6: The results on DragBench are provided in Appendix B and C (Tables 3–6). Our GoodDrag demonstrates favorable performance compared to state-of-the-art methods, consistently achieving superior results across different metrics, including DAI, GScore, and MD.

Dear Reviewer 9Jgw,

As the discussion phase is coming to a close, we wanted to check if our responses have fully addressed your concerns. If there are any remaining issues that need clarification, please let us know—we’d be happy to provide additional details.

Thank you!

Dear Reviewer 9Jgw,

Thank you for your valuable and constructive feedback on our submission. Your comments have been instrumental in improving the quality of our manuscript.

As the discussion deadline approaches, we want to ensure that all your concerns have been fully addressed. Please feel free to reach out if you require further clarifications or additional analyses.

To address your points:

• We elaborate on the unique contribution of AlDD, demonstrating that a subtle adjustment in the denoising order can have a substantial impact on drag editing — an insight that no prior work has explored or shown.

• We provide additional details on how to select the hyperparameter $B$.

• We highlight the task diversity of Drag100, showcasing how it goes beyond DragBench by encompassing a broader range of editing tasks.

• Results for DragBench with MD and IF are in Appendices B and C, with comparisons for DAI, GScore, MD, and IF detailed in Tables 1–6.

• The paper’s organization has been updated as follows: 1) A typo error “ALDD” was corrected to “AlDD”. 2) Additional space in Lines 54 have been removed as suggested.

• A new section has been added in Appendix N to discuss the limitations of our work.

We hope these updates address your concerns and reflect the improvements made to the paper. If you have had a chance to review our rebuttal and revisions, we would greatly appreciate your feedback or confirmation on whether the updated manuscript meets your expectations.

Additionally, we hope you might consider raising your score based on the revisions, as we believe these updates, guided by your thoughtful review, have substantially improved the manuscript.

Thank you once again for your time and insights.

We thank the reviewer for the valuable comments on this paper and address the raised issues here and in the revised manuscript.

Q1: Comparison with DragGAN

A1: We focus on comparisons with diffusion-based methods in Figure 7 and provide comparisons with DragGAN in Figure 8. In response to the reviewer’s suggestion, we have included DragGAN’s results for the examples in Figure 7 in Figure 13 of the revised Appendix E. Additionally, quantitative results for DragGAN are presented in Table 8 of the revised Appendix E.

It is important to note that DragGAN has significant limitations in terms of efficiency, as it requires fine-tuning the GAN generator for each new input image, resulting in a much slower process compared to our method. Due to the tight timeline of the rebuttal phase, we were only able to evaluate DragGAN on the six images in Figure 7. Nevertheless, our results demonstrate that GoodDrag achieves superior performance both qualitatively and quantitatively. We plan to conduct a more comprehensive evaluation, including additional examples and detailed performance comparisons, in the final version of the paper.

Q2: Moving training hardware and memory to main text

A2: Thanks for the suggestion. We have moved the details of the training hardware and memory usage to Section 5.1 of the revised main paper.

Q3: Moving user study to main text

A3: Thanks for the suggestion. We have moved the user study to Section 5.2 and Figure 9 of the revised main paper.

Q4: L355-357 is not needed

A4: Thank you for the suggestion. We have removed these lines as suggested.

Q5: Making figures bigger

A5: We have enlarged Figure 5, 6, and 7 for better readability.

Q6: Adding an ethics/societal impact section

A6: Following the suggestion, we have added a new section in Appendix O discussing the potential societal impact of our work.

Dear Reviewer jGsQ,

As the discussion phase is coming to a close, we wanted to check if our responses have fully addressed your concerns. If there are any remaining issues that need clarification, please let us know—we’d be happy to provide additional details.

Thank you!

We thank the reviewer for the valuable comments on this paper and address the raised issues here and in the revised manuscript.

Q1: Comparison with more recent works

A1: Thanks for pointing out the related works. We have provided comparisons against more recent methods, including DragNoise and EasyDrag, in the revised Appendix D. These comparisons are presented both qualitatively in Figure 12 and quantitatively in Table 7. The results clearly demonstrate that our proposed method achieves superior performance, which delivers more accurate drag editing, produces images with significantly higher quality, and reduces artifacts compared to the baseline methods. These related works have also been cited in the main paper.

Q2: Results on DragBench, especially with MD and IF metrics

A2: We present the results on DragBench in Appendices B and C, where the proposed GoodDrag significantly outperforms baseline approaches on both Drag100 and DragBench datasets across the DAI, GScore, and MD metrics.

Detailed comparisons for DAI and GScore on Drag100 and DragBench are provided in Table 1-4 of our paper. The MD and IF results are provided in Table 5 and 6 of the revised Appendix C, and we also include them below for easier reference:

While the IF score of our method is slightly lower than other approaches, we argue that the IF metric is fundamentally flawed as an evaluation measure for drag editing approaches. IF is defined as 1-LPIPS between the drag-edited image and the input image, meaning it penalizes any changes to the image, even when such changes are necessary to achieve the desired editing. As a result, the metric rewards outputs that are identical or nearly identical to the input image, which contradicts the very purpose of drag editing.

This inherent limitation is clearly demonstrated in Figure 11 of the revised Appendix C, where an example with superior visual quality receives the worst IF score, underscoring its failure to reflect meaningful drag edits. In contrast, the proposed GScore metric better correlates with human perception, making it a more reliable and appropriate metric for evaluating drag editing algorithms.

Dear Reviewer dZPL,

As the discussion phase is coming to a close, we wanted to check if our responses have fully addressed your concerns. If there are any remaining issues that need clarification, please let us know—we’d be happy to provide additional details.

Thank you!

Dear Reviewer dZPL,

Thank you for your valuable and constructive feedback on our submission. Your comments have been instrumental in improving the quality of our manuscript.

As the discussion deadline approaches, we want to ensure that all your concerns have been fully addressed. Please feel free to reach out if you require further clarifications or additional analyses.

To address your points:

• Comparisons with recent methods, including DragNoise and EasyDrag, are provided in Appendix D, with qualitative results in Figure 12 and quantitative results in Table 7.

• Results for DragBench with MD and IF are in Appendices B and C, with comparisons for DAI, GScore, MD, and IF detailed in Tables 1–6.

We hope these updates address your concerns and reflect the improvements made to the paper. If you have had a chance to review our rebuttal and revisions, we would greatly appreciate your feedback or confirmation on whether the updated manuscript meets your expectations.

Additionally, we hope you might consider raising your score based on the revisions, as we believe these updates, guided by your thoughtful review, have substantially improved the manuscript.

Thank you once again for your time and insights.