Research map / Image generation / Editing and style transfer
FastDrag: Manipulate Anything in One Step
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 FastDrag, a one-step drag-based image editing method on latent diffusion that replaces iterative latent optimization with a latent warpage function, bilateral nearest neighbor interpolation, and self-attention key/value consistency preservation.
Problem
Drag-based image editing with diffusion models relies on n-step iterative latent optimization (motion- or gradient-based), which is time-consuming and limits practical use.
Approach
Treats drag instructions as external forces on a stretched material to compute per-pixel warpage vectors via a latent warpage function, relocating latent pixels in one step; fills resulting null regions with bilateral nearest neighbor interpolation and preserves consistency by injecting inversion self-attention keys/values during sampling, using an LCM-distilled U-Net.
Key results
- FastDrag takes 3.12 seconds per point on an RTX 3090, nearly 700% faster than DiffEditor (21.68 s) and 2800% faster than DragDiffusion (1 min 21.54 s) on DragBench.
- FastDrag achieves drag and fidelity metrics (MD, 1-LPIPS) comparable to state-of-the-art methods on DragBench, with better qualitative drag quality reported.
- Ablations show BNNI gives best MD among null-region fill strategies and consistency-preserving strategy improves image fidelity; 10 inversion steps chosen.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace iterative latent optimization with one-step warpage optimization | FastDrag · DragBench | time per point: 3.12 seconds; nearly 700% faster than DiffEditor (21.68 seconds) and 2800% faster than DragDiffusion (1 min and 21.54 se | Helped |
| Use FastDrag instead of DragDiffusion with single-step diffusion | FastDrag · Single-step diffusion editing | editing time: 20.7 seconds -> 2.88 seconds | Helped |
| Set inversion steps to fewer than 6 (ablation) | FastDrag with LCM-equipped U-Net · DragBench | image fidelity | Hurt |
| Set inversion steps above 6 (ablation) | FastDrag with LCM-equipped U-Net · DragBench | image fidelity | Helped |
| Set inversion steps above 14 (ablation) | FastDrag with LCM-equipped U-Net · DragBench | drag effect | Hurt |
| Use 10 inversion steps (ablation) | FastDrag with LCM-equipped U-Net · DragBench | drag effect and image fidelity | Helped |
| Add bilateral nearest neighbor interpolation for null latent regions (ablation) | FastDrag · DragBench | drag precision (MD) | Helped |
| Replace BNNI with interpolation using the original value (ablation) | FastDrag · DragBench | drag effect | Hurt |
| Replace BNNI with zero-value interpolation (ablation) | FastDrag · DragBench | drag quality | Hurt |
| Replace BNNI with random-noise interpolation (ablation) | FastDrag · DragBench | drag quality | Hurt |
| Add the consistency-preserving strategy using inversion self-attention keys and values (ablation) | FastDrag · DragBench | image consistency | Helped |
| Increase the sampling step at which consistency-preserving starts (ablation) | FastDrag · DragBench | image fidelity (1-LPIPS) | Hurt |
Surprising findings
- Even a single-step diffusion setting leaves DragDiffusion at 20.7 s versus 2.88 s for FastDrag, indicating latent optimization dominates DragDiffusion's time.
Limitations
- Loses fine textures from original images (overly smooth results).
- Long-distance drags can lose object details due to lower-dimensional latent space.
- Results depend heavily on precise drag instructions (mask and handle point placement).
Experimental setup
Datasets: DragBench
Benchmarks: DragBench
Metrics: mean distance (MD) 1-LPIPS (image fidelity) time per point
Baselines: DragDiffusion DragNoise FreeDrag GoodDrag DiffEditor DragGAN StableDrag SDEDrag EasyDrag DragonDiffusion
Builds on: DragDiffusion Stable Diffusion latent consistency model DragGAN Null-text Inversion
Closest papers by meaning
- LightningDrag: Lightning Fast and Accurate Drag-based Image Editing Emerging from Videos (ICML 2025 · similarity 0.91)
- Localize, Understand, Collaborate: Semantic-Aware Dragging via Intention Reasoner (NeurIPS 2024 · similarity 0.89)
- DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models (ICLR 2024 · similarity 0.89)
- DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing (CVPR 2024 · similarity 0.89)
- CLIPDrag: Combining Text-based and Drag-based Instructions for Image Editing (ICLR 2025 · similarity 0.88)
- Drag Your Noise: Interactive Point-based Editing via Diffusion Semantic Propagation (CVPR 2024 · similarity 0.88)
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.