ThinkDrag: Semantic Drag-Based Image Editing with Visual Reasoning

Armando Fortes   Tianyi Wei   Chindanai Trakantannarong   Shangchen Zhou   Xingang Pan

S-lab, Nanyang Technological University

---

ThinkDrag interprets sparse drag handles as semantic edits

It can reason about the intended transformation before generating.

📖 Overview

ThinkDrag is a drag-based image editing method built on a unified multimodal model. Rather than treating drags as geometric displacements, it interprets them as semantic transformations and can optionally produce a visual reasoning trace before generating the edit. To support this, we construct a training dataset of semantic drag transformations paired with reasoning traces, and introduce DragBench++, a benchmark of challenging drag-based editing scenarios with reference edits.

📮 Update

• [2026.10] This repo is created.

🚧 TODO

• Release DragBench++
• Release Dataset
• Release Model Weights
• Release Inference Code

🖼️ Results

Semantic transformations. Same input and drag handles across drag-based image editing methods.

Reasoning-guided edits. ThinkDrag can first infer the intended edit from the handles, then use that reasoning trace to guide generation.

🧩 Method

Dataset. Synthetic image pairs are filtered, matched with semantic keypoints, and paired with reasoning chains.

Model. Drags are represented as structured text tokens and Drag Endpoint Markers in a unified multimodal model.

📑 Citation

If you find our work useful, please consider giving this repo a ⭐ to support our work and citing the following paper:

@article{fortes2026thinkdrag,
  title   = {ThinkDrag: Semantic Drag-Based Image Editing with Visual Reasoning},
  author  = {Fortes, Armando and Wei, Tianyi and Trakantannarong, Chindanai and Zhou, Shangchen and Pan, Xingang},
  journal = {arXiv preprint arXiv:2610.13316},
  year    = {2026}
}

©️ License

This project is licensed under NTU S-Lab License 1.0. Redistribution and use should follow this license.

🤝 Acknowledgements

We would like to thank the drag-based editing community and the many works that paved the way for this one since Drag Your GAN. In particular, we compare against DragDiffusion, DragNoise, DragLoRA, FastDrag, GeoDrag, InstantDrag & LightningDrag.

We also thank the following projects that made this work possible:

• BAGEL is the base unified multimodal model that ThinkDrag is built upon.
• Qwen3-VL is used to write source-image and edit prompts for the dataset.
• Qwen-Image, FLUX.2 & Z-Image are used to synthesize the dataset image pairs.
• Gemini 3 is used to estimate semantic keypoints and write reasoning traces.