Research map / Image generation / Editing and style transfer
GoodDrag: Towards Good Practices for Drag Editing with Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GoodDrag, which alternates drag and denoising steps within diffusion sampling plus information-preserving motion supervision, and introduces the Drag100 benchmark with DAI and Gemini-based GScore metrics for drag editing.
Problem
Diffusion-based drag editing is unstable: accumulated perturbations from performing all drag operations at one timestep cause distortions, and handle points drift away from their original features, causing artifacts and failed point control.
Approach
Distribute drag operations across diffusion timesteps by alternating B drag steps with one DDIM denoising step (AlDD), and supervise motion against the original handle-point features rather than the drifting current features (information-preserving motion supervision), with multiple supervision steps per drag to compensate for the harder optimization.
Key results
- GoodDrag achieves DAI of 0.0696 at gamma=1 versus 0.1477 for DragDiffusion, and is lowest across all patch radii.
- GoodDrag obtains an average GScore of 7.94 on Drag100, versus 6.87 for DragDiffusion and 5.38 for SDE-Drag.
- Gemini-based GScore reaches Spearman correlation 0.708 with human rankings, while TReS, MUSIQ and TOPIQ reach 0.250, -0.125 and 0.083.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GoodDrag instead of DragDiffusion | Drag editing · Drag100, γ=1 | DAI: 0.1477 -> 0.0696 | Helped |
| Use GoodDrag instead of DragDiffusion* with 210 dragging operations | Drag editing · Drag100, γ=1 | DAI: 0.1189 -> 0.0696 | Helped |
| Use GoodDrag instead of SDE-Drag | Drag editing · Drag100, γ=1 | DAI: 0.1571 -> 0.0696 | Helped |
| Use GoodDrag instead of DragDiffusion | Drag editing · Drag100 | GScore: 6.87 -> 7.94 | Helped |
| Use GoodDrag instead of SDE-Drag | Drag editing · Drag100 | GScore: 5.38 -> 7.94 | Helped |
| Use GScore instead of TReS | Image quality assessment · User study, 12 Drag100 images | Spearman correlation with human rankings: 0.250 -> 0.708 | Helped |
| Add AlDD by alternating drag and denoising operations (ablation) | GoodDrag · Drag editing, owl example | Image fidelity | Helped |
| Reduce drag operations without AlDD (ablation) | GoodDrag · Drag editing, reduced-operation example | Image fidelity | Helped |
| Reduce drag operations without AlDD (ablation) | GoodDrag · Drag editing, reduced-operation example | Drag effectiveness | Hurt |
| Add information-preserving motion supervision (ablation) | GoodDrag · Drag editing, artifact and point-control examples | Editing quality | Helped |
| Use multiple IP motion-supervision steps per drag operation (ablation) | GoodDrag with information-preserving motion supervision · Drag editing, cat example | Drag effectiveness | Helped |
| Apply IP directly without multiple motion-supervision steps per drag operation (ablation) | GoodDrag · Drag editing, cat example | Drag effectiveness | Hurt |
Surprising findings
- Naively increasing the number of drag steps for DragDiffusion (DragDiffusion*, 210 steps) only slightly improves results and still falls short of GoodDrag.
- Directly applying information-preserving motion supervision alone leads to inferior results because the loss is harder to optimize, requiring multiple supervision steps per drag.
Limitations
- Optimizing information-preserving supervision is harder, requiring multiple motion supervision steps per drag operation.
- GScore relies on a proprietary LMM (Gemini), and the user study and correlation analysis use only 12 images and 27 participants.
Experimental setup
Datasets: Drag100
Benchmarks: Drag100
Metrics: DAI (Dragging Accuracy Index) GScore (Gemini Score) user study ranking Spearman correlation with human rankings
Baselines: DragGAN DragDiffusion SDE-Drag TReS MUSIQ TOPIQ
Builds on: DragDiffusion DragGAN DDIM Latent-MasaCtrl LoRA
Closest papers by meaning
- FlowDrag: 3D-aware Drag-based Image Editing with Mesh-guided Deformation Vector Flow Fields (ICML 2025 · similarity 0.89)
- LightningDrag: Lightning Fast and Accurate Drag-based Image Editing Emerging from Videos (ICML 2025 · similarity 0.88)
- CLIPDrag: Combining Text-based and Drag-based Instructions for Image Editing (ICLR 2025 · similarity 0.88)
- FastDrag: Manipulate Anything in One Step (NeurIPS 2024 · similarity 0.88)
- Localize, Understand, Collaborate: Semantic-Aware Dragging via Intention Reasoner (NeurIPS 2024 · similarity 0.88)
- EasyDrag: Efficient Point-based Manipulation on Diffusion Models (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.