acceptodds
Withdrawn from ICLR 2027

ThinkDrag: Semantic Drag-Based Image Editing with Visual Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Drag-based image editing provides an intuitive interface for spatially controlled image manipulation by allowing users to specify handle and target points. Existing methods have made substantial progress through optimization-based, one-step, and feed-forward formulations, but they often interpret drag instructions as geometric displacements. This limits their effectiveness when the desired edit depends on interpreting image content and inferring a plausible semantic transformation from sparse point constraints. We introduce ThinkDrag, a unified multimodal framework for drag-based image editing with visual reasoning. ThinkDrag is trained to associate point constraints with meaningful object- and scene-level transformations, and can optionally produce an explicit reasoning trace that interprets the intended edit before image generation. This reasoning trace improves interpretability, particularly for ambiguous drags. Drag Endpoint Markers, motivated by an analysis of cross-image attention, provide an explicit correspondence signal that grounds each drag in image space. To support this framework, we construct a supervised dataset of semantic drag transformations paired with reasoning traces and introduce DragBench++, a benchmark targeting challenging drag-based editing scenarios with reference edit solutions. Experiments show that ThinkDrag achieves state-of-the-art performance, improving generation quality and plausibility while maintaining competitive point-following precision.

Primary area:
generative models
Keywords:
Drag-based Image Editing, Unified Multimodal Models, Visual Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MVoQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ThinkDrag proposes a unified multimodal framework for semantic drag-based image editing. It introduces a synthetic dataset of semantic drag transformations with reasoning traces and Drag Endpoint Markers (DEM) to improve spatial correspondence between drag handles and targets. Experiments on DragBench and the newly introduced DragBench++ show improved image quality and competitive point-following accuracy.

Critical strengths
  1. DEM substantially reduces point-following error compared with removing the markers (MD: 45.42 → 28.84 on DragBench), supported by attention analysis. This provides convincing evidence for the technical value of the proposed conditioning mechanism.

  2. The authors construct paired semantic edits with keypoint annotations and reasoning traces, addressing the limitations of purely geometric drag supervision. The strong overall results and user preference suggest that this data-driven approach is practically effective.

Critical weaknesses
  1. Semantic intention reasoning for drag-based editing has already been explored by prior work such as LucidDrag [1]. ThinkDrag’s main distinction is integrating reasoning and generation into a unified multimodal model and introducing DEM, but the paper does not sufficiently establish the advantages of this design over existing intention-reasoning pipelines. The absence of a direct comparison with LucidDrag makes the claimed methodological novelty difficult to assess.

  2. The gains from explicit reasoning are marginal (IF*: 0.905 → 0.911; MD: 34.24 → 33.92), with reasoning outperforming direct generation on only 51.7% of DragBench++ samples. The current ablations do not establish whether these gains arise from reasoning itself or simply from additional semantic conditioning. Furthermore, DragBench++ contains only 145 samples and uses a data-generation pipeline similar to that of the training set, including FLUX.2-generated edits. Although the source images are distinct, this setup leaves the generalization to real-world images and independently generated edits insufficiently validated. These limitations weaken the evidence for both the reasoning mechanism and the broader applicability of the method.

[1] Localize, Understand, Collaborate: Semantic-Aware Dragging via Intention Reasoner. NeurIPS 2024.

Questions for authors
  1. How does ThinkDrag differ fundamentally from LucidDrag, which also performs semantic intention reasoning for drag-based editing? Could the authors clarify the key methodological advantages of the unified-model approach?

  2. Could the authors clarify whether DEM improves performance specifically through endpoint correspondence, rather than simply adding learnable conditioning signals? Have alternative or incorrectly paired markers been tested?

  3. Since DragBench++ uses a data-generation pipeline similar to that of the training set, how do the authors ensure that the reported gains generalize beyond the synthetic editing distribution, particularly to real-world images and edits generated by different models?

Reviewer_3UMUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper adapts BAGEL to drag editing using synthetic semantic edit pairs, optional reasoning traces, and Drag Endpoint Markers. The results show improved editing quality, while the DEM ablation provides clear evidence of better spatial control.

Critical strengths
  1. The editing results are convincing, especially for changes that require coordinated object motion. ThinkDrag receives 61.0% of overall preference votes in the comparison with three baselines on DragBench.
Critical weaknesses
  1. The contribution of reasoning remains unclear. It reduces MD_MLLM from 34.24 to 33.92 and wins on only 75 of 145 examples, while increasing runtime from 14.3 to 22.9 seconds. Turning reasoning off at inference also does not isolate the benefit of reasoning supervision during training.

  2. The paper discusses semantic dragging with an intention reasoner by Cui et al. (2024), as well as LazyDrag and ContextDrag, but does not evaluate them. These methods address the ideas most relevant to ThinkDrag. Their omission leaves me uncertain whether the reported gains extend to the approaches that most directly challenge its novelty.

Questions for authors
  1. How are direct and reasoning modes mixed during training? Do you have results from the same model trained on the same image pairs without reasoning supervision?

  2. Can you provide a comparison with Cui et al. (2024), LazyDrag, and ContextDrag, or explain the constraints that prevented their evaluation? This would help establish the significance of the improvement over closely related work.

Reviewer_8MchOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this submission, the authors introduce THINKDRAG, a unified multimodal framework for drag-based image editing with visual reasoning. The motivation is that drag-editing methods primarily interpret handle-target pairs as geometric displacements, whereas more challenging edits require understanding the object structure and inferring plausible semantic transformations. The main contribution include the framework itself, the dataset of semantic drag transformations with reasoning traces, a benchmark DragBench++, and related experimental results.

Critical strengths
  1. Impressive performance. ThinkDrag achieves impressive editing quality, particularly for challenging semantic transformations. The quantitative results and user study demonstrate clear advantages over several existing methods.

  2. Effective design. DEM provides a simple and effective mechanism for spatial control, with convincing ablation results and attention analysis supporting its effectiveness.

Critical weaknesses
  1. The necessity of visual reasoning is still unclear. Semantic drag editing certainly requires visual understanding, but it is unclear why explicit reasoning is necessary. The direct editing mode already performs strongly, while reasoning provides only marginal improvements (IF*: 0.905 v.s. 0.911). Moreover, both modes share a model trained with reasoning supervision, making it difficult to isolate its contribution.

  2. Insufficient evaluation of semantic ambiguity. Sparse 2D drag constraints can correspond to multiple plausible semantic transformations. However, the user study appears to show participants only the source image, drag instructions, and candidate results, without editing prompts (Appendix B.3, Figure 12). Without knowing the intended transformation, participants can evaluate visual plausibility and point-following accuracy, but cannot reliably assess whether an edit matches the original semantic intention. Meanwhile, the reference-based IF* metric may penalize alternative valid interpretations. Given that resolving semantic ambiguity is a central motivation of the paper, this evaluation protocol does not sufficiently validate the claimed reasoning capability.

  3. The distinction from existing intention-reasoning methods, such as LucidDrag, and the relationship to geometry-aware approaches (less ambiguity with a 3D proxy), such as ObjectMorpher, need more discussion.

Reference

Localize, Understand, Collaborate: Semantic-Aware Dragging via Intention Reasoner. Neurips 2024.

ObjectMorpher: 3D-Aware Image Editing via Deformable 3DGS Models. CVPR 2026

Questions for authors
  1. What is the unique benefit of explicit reasoning? Could the authors compare full reasoning traces against concise semantic edit descriptions and a model trained without reasoning supervision?

  2. How does ThinkDrag differ from prior semantic-aware and geometry-aware drag-editing approaches? In particular, what advantages does its reasoning mechanism offer over existing intention-reasoning methods?

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

Related papers

Discussion (0)

Sign in to comment.