Drag as Evidence: Motion-Grounded Latent Recomposition for Drag-Based Editing

Xinyu Pu1,2 Hongsong Wang1,* Jie Gui1,2,* Pan Zhou3
1Southeast University 2Purple Mountain Laboratories 3Singapore Management University

NeurIPS 2026

*Corresponding authors

Paper arXiv Code (coming soon)
Editing comparison from Figure 1: point inputs, warped images, DragDiffusion, FastDrag, Inpaint4Drag, LongCat-Image-Edit, and MoRe-Drag on lion and crocodile examples
MoRe-Drag uses pixel-space warping as motion evidence to achieve accurate drag control with natural appearance.

Abstract

Modern image editors excel at semantic manipulation and visual synthesis, yet remain limited in precise spatial control, motivating the development of drag-based editing. However, existing drag-based methods often struggle to balance drag accuracy with natural, plausible, and intent-aligned generation. We propose MoRe-Drag, a motion-grounded drag-based editing method. Our key insight is to treat pixel-space warping as coarse motion evidence, and to inject this evidence into the generative sampling trajectory. Specifically, MoRe-Drag performs region-aware latent recomposition over refinement, inpainting, and anchor regions, coupled with stage-adaptive conditioning that progressively shifts from motion-grounded structure formation to semantic refinement. We further support an instruction-free interface by adapting the MLLM-based text encoder for drag-aware instruction inference. Experiments on DragBench-SR and DragBench-DR show that MoRe-Drag substantially improves drag precision over strong base editors and achieves superior drag accuracy among SOTA drag-based methods, while delivering strong semantic consistency and visually realistic results. Code and dataset will be publicly released.

Framework

MoRe-Drag framework with region-aware latent recomposition and generation-stage adaptive conditioning
Overview of MoRe-Drag. Region-aware recomposition injects motion evidence into refinement, inpainting, and preservation regions. Generation-stage adaptive conditioning uses the warped image at early timesteps to establish motion-consistent structure, then switches to the original image and the user or inferred instruction for semantic and appearance refinement.

Experiments

Quantitative Results

MethodDragBench-SRDragBench-DR
MD ↓CP ↑PF ↑MD ↓CP ↑PF ↑
DragDiffusion41.08±0.4392.75±0.3877.15±1.4931.61±0.2396.46±0.4378.45±0.95
DragNoise47.64±1.0879.00±0.2561.56±0.3130.05±0.0191.46±1.3481.02±0.18
GoodDrag28.94±0.4094.50±0.1376.50±0.3923.21±0.1295.36±0.2483.94±0.04
FastDrag23.17±0.1994.00±0.2579.62±4.3829.50±0.1292.73±0.7375.60±1.63
ContextDrag*19.0792.2582.0021.6690.9083.78
Inpaint4Drag19.79±0.1893.00±0.1381.00±0.0421.37±0.3288.53±0.2582.77±0.76
+ FLUX.1-Fill-dev18.96±0.1894.69±0.5381.41±1.4121.16±0.2891.70±0.4977.19±1.40
Qwen-Image-Edit50.48±0.1294.57±0.6383.59±0.7647.90±0.2792.43±0.4188.17±0.03
+ MoRe-Drag23.85±0.6290.15±0.1270.96±0.5826.35±0.1787.31±0.0078.75±0.89
LongCat-Image-Edit53.62±0.0994.00±0.5092.08±1.1746.73±0.2491.28±0.5591.20±0.10
+ MoRe-Drag (Ours)18.54±0.2296.36±0.3886.36±4.0620.03±0.2193.15±0.2384.82±0.98

Results from Table 1. MD: Mean Distance (↓); CP: Concept Preservation (↑); PF: Prompt Following (↑). CP and PF are scaled by 100. Results are averaged over three runs with standard deviations; * denotes results reported in the original paper. Bold indicates the best result among drag-based methods; base text editors are excluded from highlighting.

Qualitative Comparisons

Figure 6: qualitative comparisons with LongCat-Image-Edit, DragDiffusion, DragNoise, GoodDrag, FastDrag, Inpaint4Drag, and MoRe-Drag
Qualitative comparisons with state-of-the-art editing methods. Blue points denote handles; red points denote targets. Zoomed views highlight local structural fidelity.

BibTeX

@inproceedings{pu2026more,
  title={Drag as Evidence: Motion-Grounded Latent Recomposition for Drag-Based Editing},
  author={Xinyu Pu and Hongsong Wang and Jie Gui and Pan Zhou},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026},
  eprint={2609.36755},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.36755}
}