ThinkDrag: Semantic Drag-Based Image Editing with Visual Reasoning

ThinkDrag. Semantic Drag-Based Image Editing with Visual Reasoning. By Armando Fortes and four others, from S-Lab at Nanyang Technological University. Published on the ninth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Drag-based image editing gives users an intuitive handle on spatial control. They mark handle points and target points, and the image is changed to match. Existing methods have come a long way, through optimization-based, one-step and feed-forward formulations. But they often read a drag instruction as a geometric displacement, a request to move pixels from here to there. That reading falls short when the intended edit depends on understanding the image content, and on inferring a plausible semantic transformation from a few sparse points. A drag on an object might mean that it should turn, open, bend or change pose, not merely slide across the frame, and a pure displacement has no way to express that.

ThinkDrag is a unified multimodal framework that brings visual reasoning to drag-based editing. It is trained to associate point constraints with meaningful transformations at the level of objects and scenes, rather than with displacements alone. It can also, optionally, produce an explicit reasoning trace that interprets the intended edit before any image is generated. That trace makes the system more interpretable, especially when a drag is ambiguous. A user, or a researcher, can read what the system decided the drag meant before looking at the result.

Two further pieces support it. Drag Endpoint Markers, motivated by an analysis of cross-image attention, give an explicit correspondence signal that grounds each drag in image space. And because a framework like this needs suitable training data, the authors construct a supervised dataset of semantic drag transformations, each paired with a reasoning trace. For testing, they introduce DragBench++, a benchmark aimed at challenging drag-based editing scenarios, with reference edit solutions included. So the work contributes three things around the model itself: a correspondence signal, a training set that teaches the reasoning, and a benchmark that tests it.

Experiments report state-of-the-art performance for ThinkDrag. Generation quality and plausibility both improve, while precision in following the specified points stays competitive. That pairing matters here, since the risk of reading a drag semantically is drifting away from where the user actually pointed.

ThinkDrag recasts a drag as a question to be interpreted rather than a vector to be applied. The points still anchor the edit, through explicit markers, but what happens between handle and target is decided by reasoning about the content.