ThinkDrag: mmlab-ntu adds visual reasoning to drag-based image editing, claims SOTA
mmlab-ntu · hf · 2026-10-12
Researchers from mmlab-ntu introduce ThinkDrag, a unified multimodal framework for drag-based image editing with visual reasoning.
- Existing methods treat drag instructions as geometric displacements, limiting semantic edits that require interpreting image content
- ThinkDrag associates point constraints with object- and scene-level transformations and can emit an explicit reasoning trace interpreting the intended edit, improving interpretability for ambiguous drags
- Drag Endpoint Markers, motivated by cross-image attention analysis, provide explicit correspondence signals grounding drags in image space
- The team built a supervised dataset of semantic drag transformations with reasoning traces and released DragBench++, a benchmark for challenging drag-editing scenarios
- Experiments show state-of-the-art generation quality and plausibility while keeping competitive point-following precision
More from Multimodal
- Grok Imagine video generation demo showcased — creatoroff · 2026-10-12
- Imagen generates polished trading card booster pack artwork — Dimillian · 2026-10-12
- MiniMax Design builds a scroll-exploding watch landing page in one prompt — sujingshen · 2026-10-12
- 30 art styles, 8 local image models on one Mac via mflux: a same-seed comparison — redditstig · 2026-10-12
- 'The Last Harvest': AI-generated vision of a civilization mining the last black holes — burny_tech · 2026-10-12
- Three Kingdoms AI short film 'The Year of Jiazi' released with English subtitles — Unusual-Leek-2116 · 2026-10-12