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.
Figures & tables
Figure 1: ThinkDrag handles challenging and ambiguous drags, optionally making its interpretation explicit as a reasoning trace that the user can refine, while the drags keep the edit spatially grounded.
Figure 2: Dataset construction pipeline of the ThinkDrag dataset.
Figure 3: Diverse examples of the ThinkDrag dataset.
Figure 4: Overview of the ThinkDrag unified multimodal model pipeline.
DragBench++
DragBench
Method
Type
IF ↑
IF* ↑
MD ↓
MD MLLM↓
IF ↑
FID ↓
MD ↓
MD MLLM↓
Reference
–
–
43.42
27.38
–
–
–
–
DragDiffusion
TTO
0.833
0.819
71.69
–
0.882
38.90
30.88
–
DragNoise
0.867
0.829
74.44
–
0.900
44.39
32.70
–
DragLoRA
0.808
0.810
48.36
59.98
0.874
43.35
23.20
36.54
FastDrag
OS
0.780
0.771
79.02
–
0.859
49.34
27.45
–
Table 1: Results on both benchmarks. TTO: test-time optimization, OS: one-step, FT: fine-tuned.
Table 6
Figure 5: Qualitative comparison with other models on DragBench and DragBench++.
Figure 6: Qualitative comparison between direct generation and the reasoning mode on DragBench++.
Figure 7: Joint text-drag conditioning. Given the same drag instruction, different text prompts guide the model toward different valid edits, while the drags keep each edit spatially grounded.
Figure 10
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Additional visual comparison with other models on DragBench and DragBench++. Target points are overlaid on the generated images to facilitate analysis of editing accuracy.
Number of drags
1
2
3
≥ 4
All
Samples
28
51
37
29
145
Reasoning wins
10
25
24
16
75
Win rate
35.7%
49.0%
64.9%
55.2%
51.7%
Appendix
Table 6: Win rate of reasoning over direct generation on DragBench++ by drag number (MD MLLM ).
Edit type
Open
Adjust
Move
Rotate
Raise
Other
Samples
20
23
20
19
19
44
Win rate
60.0%
52.2%
50.0%
47.4%
42.1%
54.5%
Appendix
Table 7: Win rate of reasoning over direct generation on DragBench++ by edit type, determined by the first verb of the edit instruction. Verbs with fewer than five samples are grouped under Other.
Model
Correct
Accuracy
BAGEL
95
38.0%
ThinkDrag
182
72.8%
Appendix
Table 8: Coordinate grounding on 250 annotated handle points from DragBench++.
Figure 10: Grounding of textual coordinates. Each input image shows the queried normalized coordinate as a red point. Given the image and the coordinate, BAGEL often describes an unrelated object, whereas ThinkDrag identifies the correct object and part at the queried location.
Benchmark
Drags
w/o DEM
w/ DEM
DragBench++
294
6.1%
50.7%
DragBench
226
13.3%
46.0%
Appendix
Table 9: Fraction of drags whose cross-image attention from the target point peaks at the handle (within one latent cell), with and without DEM.
Method
Runtime (s)
Peak Memory (GB)
ThinkDrag
14.3
35.5
ThinkDrag w/ Reasoning
22.9
35.6
Appendix
Table 10: Runtime and peak memory usage.
Figure 11: Comparison between DIFT-based and MLLM-based point tracking. MLLM-based tracking is more robust in challenging cases involving large structural changes or distractors.
Benchmark
Method
Edit accuracy
Generation quality
Overall preference
DragBench++
DragLoRA
7.7
6.3
6.3
GeoDrag
7.0
6.7
6.3
LightningDrag
7.7
8.3
7.7
ThinkDrag
77.7
78.7
79.7
DragBench
DragLoRA
13.7
15.7
13.3
GeoDrag
7.7
8.0
8.7
Appendix
Table 11: User study vote share (%) by benchmark and criterion. Each benchmark contributes 10 of the 20 questions.
Figure 12: Example question from the user study interface.
Stage
Model
Output and key details
Source prompts
Qwen3-VL
One-sentence scene prompt of 45–110 words from a subject and a location drawn from a pool. The subject is dominant, object parts appear with their parent object, and 8–14 location details vary the scene.
Source images
Qwen-Image, FLUX.2, Z-Image
One image per prompt, square in about 70% of cases and otherwise landscape or portrait, as chosen by the prompt model.
Edit proposal
Qwen3-VL
Five edit prompts per source image, grounded in visible parts and achievable with a few drags: articulation ( e.g. , opening a door, raising a limb), repositioning of an object, or deformation of a boundary. Insertion, removal, and style, lighting, or material changes are excluded.
Target images
FLUX.2
One edited image per source image and edit prompt.
Filtering
Qwen3-VL
Binary decision on each pair. A pair is kept only if the edit is clearly achieved in the intended direction, identity and unrelated content are preserved, the change is local, and there are no artifacts, anatomy or mechanics violations, or global color and lighting shifts.
Keypoints
Gemini 3 Pro
One to seven keypoints on moved and anchor parts, located in both images on a 0–1000 grid, preferring corners, tips, and joints. Pairs judged unsuitable receive no keypoints and are discarded, as are drags shorter than 32 grid units.
Appendix
Table 12: Stages of the dataset construction pipeline.
Beijing University of Posts and Telecommunications · Beijing Key Laboratory of Multimodal Data Intelligent Perception and Governance · Tencent Hunyuan +1