Research map / Image generation / Editing and style transfer
EasyDrag: Efficient Point-based Manipulation on 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.
EasyDrag performs point-based image dragging on Stable Diffusion 1.5 without LoRA fine-tuning or user masks, using stable motion supervision, automatic gradient-based masks, and reference guidance during denoising.
Problem
Point-based image editing with diffusion models is slow and user-unfriendly because DragDiffusion needs per-image LoRA fine-tuning and hand-drawn masks, and GAN-based DragGAN is limited by StyleGAN's generative domain.
Approach
Drag handle points by optimizing a DDIM-inversion latent using cosine feature similarity to original-image U-Net features (always sampled from the original features, not drifting ones), derive masks automatically from normalized latent gradients, and preserve identity by mixing in reference-guided denoising with DDIM-inversion latents.
Key results
- On DragBench, EasyDrag (w/ mask) reaches IF 0.889 and MD 29.07 versus DragDiffusion's IF 0.882 and MD 36.71; the mask-free version gets IF 0.910 and MD 32.55.
- EasyDrag runs in about 30.6 s per edit without LoRA fine-tuning, versus 69.4 s for DragDiffusion on a single A100.
- Replacing DragDiffusion's point-tracking motion supervision with the stable version yields MD 29.07 vs 52.21 for DragonDiffusion's supervision (ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EasyDrag instead of DragDiffusion | point-based image editing on Stable Diffusion 1.5 · DragBench | IF: 0.882 -> 0.910 | Helped |
| use EasyDrag instead of DragDiffusion | point-based image editing on Stable Diffusion 1.5 · DragBench | MD: 36.71 -> 32.55 | Helped |
| use EasyDrag instead of DragDiffusion | point-based image editing on Stable Diffusion 1.5 · DragBench | Time (s): 69.4 -> 30.6 | Helped |
| remove LoRA fine-tuning (ablation) | DragDiffusion · DragBench | MD: 36.71 -> 53.63 | Hurt |
| remove LoRA fine-tuning (ablation) | DragDiffusion · DragBench | Time (s): 69.4 -> 33.7 | Helped |
| remove the mask (ablation) | DragDiffusion · DragBench | IF: 0.882 -> 0.766 | Hurt |
| use a user-provided mask instead of auto mask generation (ablation) | EasyDrag · DragBench | IF: 0.910 -> 0.889 | Hurt |
| add stable motion supervision instead of DragGAN motion supervision | point-based image editing on Stable Diffusion 1.5 · DragBench | MD: 51.32 -> 29.07 | Helped |
| add stable motion supervision instead of DragonDiffusion motion supervision | point-based image editing on Stable Diffusion 1.5 · DragBench | MD: 52.21 -> 29.07 | Helped |
| use concatenated feature maps from upper layers 1&2 (ablation) | stable motion supervision · DragBench | performance | Helped |
| set motion step K to 12 (ablation) | EasyDrag · DragBench | performance | Helped |
Surprising findings
- The mask-free variant achieves higher image fidelity (0.910) than the mask-guided variant (0.889), though with slightly worse mean distance.
Limitations
- Small editing regions or long dragging distances give unsatisfactory results.
- Self-attention can cause unintended changes in non-edited areas for generated images.
- Out-of-distribution images may not reconstruct well without reference guidance.
Experimental setup
Datasets: DragBench
Benchmarks: DragBench
Metrics: Image Fidelity (IF) Mean Distance (MD) runtime (s)
Baselines: DragGAN DragDiffusion DragDiffusion w/o LoRA DragDiffusion w/o mask SDE-Drag DragonDiffusion
Builds on: DragGAN DragDiffusion DragonDiffusion DIFT DDIM inversion
Closest papers by meaning
- DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing (CVPR 2024 · similarity 0.91)
- FastDrag: Manipulate Anything in One Step (NeurIPS 2024 · similarity 0.88)
- LightningDrag: Lightning Fast and Accurate Drag-based Image Editing Emerging from Videos (ICML 2025 · similarity 0.88)
- GoodDrag: Towards Good Practices for Drag Editing with Diffusion Models (ICLR 2025 · similarity 0.87)
- 3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing (ICLR 2025 · similarity 0.87)
- CLIPDrag: Combining Text-based and Drag-based Instructions for Image Editing (ICLR 2025 · similarity 0.86)
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.