- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes CannyDrag, a DBIE framework that incorporates Canny edge guidance to address the ambiguity of point-based dragging. The method progressively manipulates Canny maps to support point-level deformation, line-level curve bending, and region-level geometric transformations. Experiments on DragBench and two newly constructed benchmarks demonstrate improved editing quality.
Critical strengths
-
The paper is well written and provides detailed explanations of the motivation and the method. The pipeline is clear and easy to follow. The results also demonstrate noticeable improvements over existing methods, especially in spatial editing clarity and accuracy.
-
The paper makes good use of the ControlNet by transforming sparse drag instructions into explicit Canny edge guidance. This provides more direct control over the intended geometric changes and enables a unified framework for 3 types of changes. I find this design intuitive and practically useful.
Critical weaknesses
-
The ambiguity of point-based dragging as already been discussed in DragNeXt. Although CannyDrag introduces a different approach based on Canny guidance, the paper does not discuss or compare with DragNeXt, making it difficult to assess the significance of its technical contributions over this closely related work.
-
Several important experimental details are missing: end-to-end runtime givem per-image fine-tuning; how baselines are adapted and evaluated fairly on the new tasks, whether they receive the same editing information; doesn't isolate the effect of progressive Canny guidance over static guidance. Taken together, these issues make the experiment part less comprehensive.
Questions for authors
-
Can the authors provide a comparison with DragNeXt and a systematic analysis of why CannyDrag performs better?
-
Does the per-image LoRA fine-tuning significantly affect the practicality of the method? Can the authors provide runtime and cost comparisons to evaluate this issue?
-
How are the baselines evaluated fairly on the new tasks?
-
What is the motivation for using progressive Canny guidance instead of static guidance? Can the authors provide a component ablation to demonstrate its benefits?
2 comments
3. Fair Comparison on New Tasks (Weakness 2; Question 3)
We appreciate the reviewer's concern regarding fair comparisons across methods with different editing interfaces. All methods share the same source images, text prompts, editable masks, editing objectives, and evaluation metrics. For DragBench-Line and DragBench-Region, we additionally introduce line_area and region_area to specify the structures to be manipulated, while the editable mask defines where changes are allowed.
Different methods support different input interfaces. RegionDrag can directly utilize the annotated region and original control points through its native Region + points interface. In contrast, point-based baselines cannot directly accept line- or region-level annotations. We therefore convert the annotated structures into supplementary handle-target pairs to convey the intended geometric transformations through their existing interfaces.
Specifically, we sample five pairs along each line centerline, including two fixed endpoints, and ten pairs within each region using farthest-point sampling. Region targets follow the annotated similarity transformation. These pairs broadly cover the manipulated structures and serve as shared evaluation locations for all methods. CannyDrag uses the same pairs for motion supervision and point tracking, while using line_area and region_area to construct task-specific canny conditions.
Thus, all methods share consistent editing objectives and are evaluated at identical geometric correspondences, while their editing inputs follow their respective interfaces.
Additional comparison with RegionDrag. To further evaluate CannyDrag against region-based methods that directly accept region_area inputs, we additionally compare it with RegionDrag on DragBench-Region, using its native region-editing interface and the same evaluation metrics.
| Method | MD ↓ | DAI_1 ↓ | DAI_10 ↓ | DAI_20 ↓ | IF ↑ |
|---|---|---|---|---|---|
| RegionDrag | 37.01 | 12.15 | 12.65 | 13.81 | 93.09 |
| CannyDrag | 15.28 | 6.59 | 8.41 | 9.08 | 84.71 |
CannyDrag achieves substantially lower MD and DAI across all spatial accuracy metrics, even compared with RegionDrag, which directly accepts region-level inputs. Although RegionDrag achieves higher IF, the results demonstrate the effectiveness of CannyDrag in accurately performing the intended geometric transformations.
4. Progressive Canny Guidance and Ablations (Weakness 2; Question 4)
We appreciate this question, as it relates directly to the motivation behind our progressive guidance design.
Motivation. During denoising, image content is progressively refined over multiple steps. However, static guidance imposes the final target geometry from the beginning, potentially introducing abrupt structural changes and image distortion. To address this, progressive Canny guidance gradually transforms the Canny map from the source to the target, allowing geometric constraints to evolve alongside the denoising process. This encourages smoother structural transitions rather than forcing the entire transformation at once.
Ablation study. To evaluate this design, we conduct additional controlled ablations on Canny conditioning, Canny-conditioned inversion, and progressive guidance. In particular, we compare progressive guidance with static guidance, where the fixed final canny map is applied throughout the guidance interval, while keeping the remaining editing settings unchanged.
| canny condition | Conditioned inversion | Progressive guidance | MD ↓ | DAI_1 ↓ | DAI_10 ↓ | DAI_20 ↓ | IF ↑ |
|---|---|---|---|---|---|---|---|
| ✗ | ✗ | ✗ | 21.27 | 12.55 | 11.16 | 10.40 | 86.26 |
| ✓ | ✗ | ✗ | 21.02 | 10.85 | 9.55 | 8.95 | 87.10 |
| ✓ | ✓ | ✗ | 20.70 | 10.45 | 8.93 | 8.85 | 86.96 |
| ✓ | ✓ | ✓ | 19.95 | 9.47 | 8.80 | 8.77 | 86.32 |
The results show that progressive guidance consistently outperforms static guidance across all spatial accuracy metrics, supporting the benefit of gradually introducing geometric constraints. Canny-conditioned inversion provides further improvements. Overall, the complete method achieves the best MD and DAI scores, while maintaining comparable image fidelity.
We sincerely appreciate the reviewer's thoughtful comments and constructive suggestions. These questions provide a valuable opportunity to clarify our methodology and strengthen the experimental evaluation. Below, we respond to each concern and question point by point.
1. Comparison with DragNeXt (Weakness 1; Question 1)
We appreciate the reviewer highlighting DragNeXt's earlier discussion of drag ambiguity, including transformation types and manipulated regions. CannyDrag further explores multi-granularity geometric manipulation and explicit structural guidance, with three key distinctions:
(1) Explicit, high-resolution structural conditioning. DragNeXt uses geometric correspondences for latent feature optimization. In contrast, CannyDrag transforms geometric instructions into explicit Canny conditions at the original image resolution. This enables fine-grained control over object boundaries and thin structures.
(2) Multi-granularity geometric manipulation. CannyDrag unifies point-, line-, and region-level editing through task-specific geometric mappings. In particular, line-level editing uses spline interpolation to guide continuous curve deformation, which cannot be fully specified by sparse control points. Region-level editing also supports uniform scaling.
(3) Progressive guidance and interactivity. CannyDrag uses the original canny map during inversion and progressively transformed canny maps during denoising. This provides structural guidance throughout editing. Moreover, Canny maps offer interpretable previews, allowing users to inspect and refine the target geometry before generating the edited image.
We further conduct direct quantitative comparisons with DragNeXt on all three benchmarks, as summarized below.
CannyDrag consistently achieves lower MD and DAI across all three benchmarks, demonstrating more accurate spatial manipulation. While DragNeXt achieves higher IF scores, qualitative results show that it sometimes preserves the original image without sufficiently performing the intended edits. Since IF measures similarity to the source image, such behavior can lead to higher IF scores despite lower geometric accuracy. We have included the corresponding quantitative results and qualitative comparisons in the newly uploaded revised PDF.
2. Computational Cost and Efficiency (Weakness 2; Question 2)
We appreciate this practical consideration, as computational efficiency is an important factor in interactive image editing. Per-image LoRA adaptation is also adopted by existing drag-editing methods, such as DragDiffusion and GoodDrag. Therefore, it is not an additional requirement unique to CannyDrag. To quantify the computational cost of CannyDrag, we compare it with GoodDrag on an NVIDIA H100 GPU, using the same 70-step LoRA adaptation and 7-step dragging settings. We report end-to-end runtime, including LoRA adaptation, inversion, and editing, together with peak GPU memory usage.
The results show that CannyDrag requires more runtime and peak GPU memory under these settings, with additional computation associated with ControlNet and canny conditioning. These costs accompany the improvements in spatial editing accuracy reported in our comparisons. Moreover, per-image LoRA parameters can be reused for multiple editing instructions on the same image, reducing the average cost of subsequent edits.
To further assess the role of per-image adaptation, we evaluate CannyDrag without LoRA fine-tuning on DragBench, while retaining Canny conditioning and the same dragging settings.
The results indicate that per-image adaptation improves geometric editing accuracy and image fidelity under the evaluated settings. Together, these measurements characterize the trade-off between editing performance and computational cost, with LoRA adaptation reusable across edits of the same image.