acceptodds
Under review as a conference paper at ICLR 2027

CannyDrag: Unified Point, Line, and Region Drag Editing via Canny Guidance

Abstract

Drag-based image editing offers an intuitive method to control spatial changes through handle-target point pairs, but sparse point instructions often lead to ambiguous editing intentions. The same point movement may correspond to local deformation, line bending, or region-level transformation, while existing motion supervision mainly constrains local neighborhoods around handle points. To address this limitation, we propose CannyDrag, a canny-guided drag editing framework that extends point-based control to point-, line-, and region-level manipulation. CannyDrag first introduces canny guidance during inversion to keep the inverted latent compatible with the later denoising trajectory. It then manipulates the canny map according to the selected task type and injects the transformed canny condition progressively during denoising. This design converts sparse drag instructions into task-specific structural guidance and helps align the editing process with user intentions. Experiments on point-, line-, and region-level drag tasks demonstrate that CannyDrag improves spatial accuracy while maintaining competitive image fidelity.

Primary area:
generative models
Keywords:
image editing, diffusion model

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_arueOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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
  1. Can the authors provide a comparison with DragNeXt and a systematic analysis of why CannyDrag performs better?

  2. 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?

  3. How are the baselines evaluated fairly on the new tasks?

  4. 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

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.

BenchmarkMethodMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
DragBenchDragNeXt38.9720.5319.0618.3190.14
CannyDrag19.959.478.808.7786.32
DragBench-LineDragNeXt9.6427.0124.7319.7296.16
CannyDrag8.0721.2014.9311.7794.46
DragBench-RegionDragNeXt43.8120.1420.8920.9991.63
CannyDrag15.286.598.419.0884.71

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.

MethodLoRA (s)Total (s)Memory (GiB)
GoodDrag8.8926.9411.65
CannyDrag11.0836.9715.63

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.

SettingMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
CannyDrag w/o LoRA24.3113.0610.389.7485.05
CannyDrag w/ LoRA19.959.478.808.7786.32

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.

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.

MethodMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
RegionDrag37.0112.1512.6513.8193.09
CannyDrag15.286.598.419.0884.71

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 conditionConditioned inversionProgressive guidanceMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
✗✗✗21.2712.5511.1610.4086.26
✓✗✗21.0210.859.558.9587.10
✓✓✗20.7010.458.938.8586.96
✓✓✓19.959.478.808.7786.32

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.

Reviewer_bGsEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CannyDrag, a framework that aims to reduce ambiguity in drag-based image editing, where sparse point-pair instructions may correspond to different intended edits. It explicitly distinguishes point-, line-, and region-level operations and transforms the source Canny map through local warping, curve bending, or similarity transformations to provide task-specific structural guidance.

Critical strengths
  • The paper addresses a fundamental challenge in drag-based image editing: sparse point-pair instructions can correspond to different intended edits. Resolving this ambiguity is crucial for reliable control and alignment with user intent. Explicitly distinguishing editing types and providing corresponding structural guidance is therefore a meaningful direction.
Critical weaknesses
  • Missing comparison with key related work. The recent work [a] identifies and analyzes the ambiguity of drag-based image editing (drag type and optimization region) and proposes to address it by using geometric transformations, which is highly related to this work. However, the paper neither discusses its methodological relationship to [a] nor provides an experimental comparison. For example, the ambiguity analysis in Section 3.1 shares substantial conceptual similarities with Proposition 1 of [a], yet the paper neither cites this prior work nor explains how its analysis differs.

    1. Please explicitly discuss the relationship to [a] in three respects: (i) How does the ambiguity analysis in this paper differ from that in [a] (drag type and region)? (ii) What are the differences between the proposed geometric operations and those used in [a]? (iii) What advantages does applying geometric transformations to Canny maps offer over using geometric transformations to impose feature constraints for latent optimization, as in [a]?

    2. Given the close relevance of [a], a direct comparison is necessary to establish the advantages of the proposed method. Please include [a] in the quantitative comparisons in Tables 1 and 2 and the qualitative comparison in Figure 5.

    [a] DragNeXt: Rethinking Drag-Based Image Editing. AAAI, 2026.

  • The benefits of Canny-conditioned inversion and progressive guidance are not isolated. Figure 7 studies inversion strength and dragging duration, but it does not isolate Canny-conditioned inversion or progressive guidance from simpler alternatives. For point-level editing, Appendix C ties Canny manipulation and latent optimization to the same step count, so the sweep does not separate their effects. Consequently, the results support the complete pipeline but do not establish the specific benefits of the proposed inversion and progressive-conditioning mechanisms. Controls that remove Canny conditioning only during inversion or use a fixed final Canny map during the guidance interval would help establish these benefits. Per-image adaptation and the latent-optimization schedule and budget should remain fixed.

  • The comparison protocol for the new tasks is insufficiently illustrated. Appendix C distinguishes the additional line/region masks from the editable mask, but does not explain how the compared methods receive or use the additional structure and task information. Without this information, it is unclear how consistently the intended edits are conveyed across methods. In addition to the missing comparison with [a], Table 2 also omits RegionDrag, which is directly relevant to the region-editing setting. Although the auxiliary evaluation points are useful, the numbers of images and editing cases in the new benchmarks are not reported. These omissions make it difficult to assess benchmark coverage and reproduce the comparisons on the two new tasks.

Questions for authors
  • How does the ambiguity analysis in Section 3.1 differ from Proposition 1 of [a], particularly regarding drag type and the region to be manipulated? How do the proposed geometric operations differ from those in [a], and what additional benefits does transforming Could the authors substantiate these benefits by including [a] in the quantitative comparisons in Tables 1 and 2 and the qualitative comparison in Figure 5?

  • Could the authors separately evaluate removing Canny conditioning during inversion and replacing progressive guidance with the final transformed Canny map throughout the guidance interval? Please keep the backbone, per-image adaptation, and latent-optimization schedule and budget fixed across these comparisons. These controls would clarify whether the proposed conditioning schedule contributes beyond adding structural guidance to an existing dragging pipeline.

  • What distinct advantages does applying geometric transformations to Canny maps offer over applying them in feature space, as in [a]?

4 comments
AuthorsOpenReview ↗

Initial response (1/3)

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.

Comparison with DragNeXt (Weakness 1; Questions 1 and 3)

We appreciate the reviewer highlighting DragNeXt, as clarifying its relationship to CannyDrag is important for understanding our contributions.

  • 1.(i) (Ambiguity Analysis)

    DragNeXt identifies two sources of ambiguity in drag-based editing: how to drag (transformation type) and what to drag (manipulated region). Our analysis of instruction-level ambiguity shares this observation and further examines two aspects. First, we distinguish three levels of drag manipulation: point-level local deformation, line-level curve bending, and region-level similarity transformation. In particular, line-level editing involves the coordinated deformation of continuous structures, where sparse point displacements cannot fully specify the desired curve geometry between control points. This task is not explicitly formulated in DragNeXt, which primarily focuses on region-based geometric transformations. Second, we highlight supervision-level incompleteness. Even when the intended transformation type and affected region are specified, conventional motion supervision primarily enforces feature correspondence within local neighborhoods around handle points, leaving the surrounding structures insufficiently constrained. Therefore, while DragNeXt focuses on resolving ambiguity in transformation types and manipulated regions, CannyDrag further considers the granularity of structural manipulation and the completeness of structural supervision.

  • 1.(ii) (Geometric Operation)

    Although both methods employ geometric transformations, their formulations and supported operations differ. DragNeXt primarily uses region-level translation and 2D rotation. In contrast, CannyDrag explicitly formulates three task-specific geometric mappings: Point-level local deformation: A spatially varying displacement field is constructed using handle-target correspondences and distance-based weights, with attenuation determined by the actual editable-mask boundary. Line-level curve bending: An ordered centerline and a spline-interpolated displacement field explicitly model continuous curve deformation while preserving the relative offsets of neighboring edge pixels. Region-level similarity transformation: A partial affine transformation supports translation, in-plane rotation, and uniform scaling, with scaling not explicitly supported by the current geometric operators in DragNeXt. Furthermore, CannyDrag applies these transformations directly to binary canny edge coordinates and reconnects adjacent edge pixels to maintain structural continuity. This differs from DragNeXt, which employs geometric correspondences to construct feature-level supervision for latent optimization.

  • 1.(iii) (Advantages of Canny-based Geometric Conditioning)

    DragNeXt establishes region-level feature correspondences and optimizes latent representations. In contrast, CannyDrag transforms geometric instructions into explicit canny conditions, which guide inversion and progressive denoising through ControlNet. This design offers three main advantages: (1) High-resolution structural precision: canny maps preserve fine-grained edge information at the original image resolution, enabling more precise geometric manipulation of object boundaries and thin structures. In contrast, feature-level constraints operate on spatially compressed representations, which may reduce the precision of geometric alignment. (2) Explicit and progressive structural guidance: CannyDrag provides spatially distributed constraints for continuous curve bending and coherent region transformations, while consistently incorporating the original and progressively transformed canny conditions during inversion and denoising. (3) Enhanced user interactivity: By using canny maps as an interpretable intermediate representation, CannyDrag allows users to visualize the manipulated canny map before completing the diffusion generation process and interactively refine the editing instructions or geometric transformations without repeatedly generating full images.

AuthorsOpenReview ↗

Initial response (2/3)

2. Direct comparison. We have conducted additional quantitative and qualitative comparisons with DragNeXt on DragBench, DragBench-Line, and DragBench-Region. The quantitative results are summarized below. CannyDrag achieves lower MD and DAI across all three benchmarks. 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. Furthermore, we have updated Tables 1 and 2 to include DragNeXt and revised Figure 5 to provide direct qualitative comparisons in the newly submitted revised manuscript.

BenchmarkMethodMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
DragBenchDragNeXt38.9720.5319.0618.3190.14
CannyDrag19.959.478.808.7786.32
DragBench-LineDragNeXt9.6427.0124.7319.7296.16
CannyDrag8.0721.2014.9311.7794.46
DragBench-RegionDragNeXt43.8120.1420.8920.9991.63
CannyDrag15.286.598.419.0884.71

Ablation on conditioned inversion and progressive guidance (Weakness 2; Question 2)

The suggested controls are helpful for distinguishing the contributions of conditioned inversion and progressive guidance. To assess these contributions, we conduct additional controlled ablation experiments to evaluate canny-conditioned inversion and progressive canny guidance. Specifically, we compare variants that remove canny conditioning only during inversion or replace progressive guidance with a fixed final canny map during denoising.

canny conditionConditioned inversionProgressive guidanceMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
✗✗✗21.2712.5511.1610.4086.26
✓✗✗21.0210.859.558.9587.10
✓✓✗20.7010.458.938.8586.96
✓✓✓19.959.478.808.7786.32

The results demonstrate the incremental benefits of Canny-conditioned inversion and progressive guidance under controlled settings. Compared with the variant without canny-conditioned inversion, introducing canny guidance during inversion consistently improves MD and DAI, supporting its effectiveness in aligning inversion with conditional denoising. Furthermore, replacing fixed canny guidance with progressive guidance yields additional improvements across all spatial accuracy metrics, demonstrating the benefit of gradually introducing geometric transformations. The complete method achieves the best spatial editing accuracy while maintaining competitive image fidelity.

AuthorsOpenReview ↗

Initial response (3/3)

Comparison Protocol and Benchmark Coverage (Weakness 3)

We appreciate the opportunity to clarify how editing instructions are adapted across methods with different input interfaces. We provide additional clarification on how task-specific annotations are adapted to different baseline methods, supplement the comparison with RegionDrag, and report the detailed benchmark statistics.

1. Comparison protocols and fairness.

In DragBench-Line and DragBench-Region, we introduce line_area and region_area, respectively, to specify the area to be manipulated, whereas the editable mask specifies where changes are allowed. We introduce supplementary handle-target pairs for two reasons. First, they provide the same evaluation locations for all methods, enabling a consistent comparison over the manipulated structure. Second, point-based methods do not directly support line/region task-area inputs, so they use these supplementary pairs to perform drag editing through their existing point interfaces. CannyDrag uses the same supplementary pairs for motion supervision and point tracking for fairness. Specifically, we sample five pairs along each line centerline, including two fixed endpoints, and ten pairs within each region using farthest-point sampling, with region targets obtained from the annotated similarity transformation. These points are distributed to cover the annotated line or region as broadly as possible.

CannyDrag additionally uses line_area for curve bending and region_area for similarity transformations to construct canny conditions. RegionDrag also receives the annotated region and original control pairs through its official Region + points preprocessing. The comparison therefore respects each method's supported interface while sharing editing objectives and evaluation locations.

2. Additional comparison with RegionDrag.

We evaluate RegionDrag on DragBench-Region cases using its official editing implementation.

MethodMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
RegionDrag37.0112.1512.6513.8193.09
CannyDrag15.286.598.419.0884.71

CannyDrag achieves lower MD and DAI than RegionDrag, even though both methods use the annotated region, while RegionDrag obtains higher IF. This comparison supports CannyDrag's spatial editing accuracy against an area-aware baseline, rather than only methods that lack region inputs.

3. Benchmark coverage and reproducibility.

All source images in DragBench-Line and DragBench-Region come from DragBench and are reannotated for line-level and region-level editing, respectively. Together, the two benchmarks cover all categories in the original DragBench. Each benchmark contains 100 source images. We will expand Appendix C to provide the benchmark statistics, annotation definitions and inference settings.

Thanks for the authors' response. My main concerns have been addressed and I will keep my positive score.

Reviewer_JXB1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CannyDrag for Canny-guided drag-based image editing. CannyDrag transforms point-pair instructions into Canny maps for point-level deformation and region-level similarity transformation. Canny maps can also be incorporated into inversion and progressive denoising to provide structural guidance. Experiments across the three tasks demonstrate the effectiveness of this method.

Critical strengths
  1. The motivation is clear; the writing is easy to follow.
  2. The method is reasonable and effective. CannyDrag achieves the best MD and DAI across all three evaluated tasks.
Critical weaknesses
  1. The ablations are limited. Only inversion strength and dragging steps are discussed. Could the authors provide more ablation experiments to validate the effectiveness of the Canny-related modules?

  2. The U-Net baseline SD1.5 is somewhat outdated. Could this method be extended to DiT based models (SD3, FLUX)? If so, Could the authors provide the related results of CannyDrag?

Questions for authors

See Weaknesses.

1 comment

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.

canny-related component ablations (Weakness 1)

We appreciate the suggestion to examine the individual contributions of the Canny-related components. To assess these contributions, we conduct additional ablation experiments on canny conditioning, canny-conditioned inversion, and progressive canny guidance. The results show that each component progressively improves spatial editing accuracy. The canny condition provides explicit structural constraints, conditioned inversion improves consistency with canny-guided denoising, and progressive guidance further refines geometric alignment. The full method achieves the best results across all spatial accuracy metrics while maintaining competitive image fidelity.

canny conditionConditioned inversionProgressive guidanceMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
✗✗✗21.2712.5511.1610.4086.26
✓✗✗21.0210.859.558.9587.10
✓✓✗20.7010.458.938.8586.96
✓✓✓19.959.478.808.7786.32

Extension to DiT backbones (Weakness 2)

This is an important direction for extending our approach to DiT backbones. To explore whether our canny-guidance mechanism can also benefit DiT-based editing, we build upon the FLUX-based DragFlow (ICLR 2026) framework, which uses region-level geometric supervision for drag-based editing. Specifically, we integrate our proposed canny-guidance mechanism into DragFlow by introducing a frozen canny ControlNet, using the original canny map during inversion and progressively transformed canny maps during editing, while keeping the remaining DragFlow settings unchanged. We compare the original DragFlow with DragFlow+canny guidance on DragBench-Region.

MethodMD ↓DAI_1 ↓DAI_10 ↓DAI_20 ↓IF ↑
DragFlow46.5213.4314.6514.8890.55
DragFlow+canny44.7813.5114.5614.8290.76

The results show that incorporating canny guidance improves MD, DAI_10, DAI_20, and IF while maintaining a comparable DAI_1 score. These results suggest that explicit canny-based structural conditioning can complement region-level feature supervision in the FLUX-based DragFlow framework, supporting the transferability of our canny-conditioning mechanism to this DiT-based editing setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.