Research map / Image generation / Editing and style transfer
FreeDrag: Feature Dragging for Reliable Point-based Image Editing
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 FreeDrag, a feature-dragging method for point-based image editing that uses adaptive template features and line search with backtracking to fix miss tracking and ambiguous tracking, and introduces FreeDragBench with a CCSD metric.
Problem
Point-based drag editing (e.g., DragGAN, DragDiffusion) suffers from miss tracking of handle points and ambiguous tracking where points latch onto similar-looking regions, causing distorted or inaccurate edits.
Approach
Replace point tracking with feature dragging: each handle point is pulled toward an exponentially updated template feature whose update scale adapts to drag quality, and the handle point is constrained to a line toward its target with a line search and backtracking to discard abnormal moves.
Key results
- On FreeDragBench, FreeDrag lowers CCSD across all categories versus DragGAN (e.g., Face 0.35 vs 0.83) and FID for Face (29.50 vs 38.07).
- On DragBench, FreeDrag reduces mean distance from 38.76 (DragDiffusion) to 33.49 and LPIPS (x10) from 1.38 to 1.02, with runtime 63.62s vs 71.77s.
- Face-model ablation shows CCSD of 0.82 without adaptive updating, 0.52 without backtracking, and 0.35 for the full method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Face | FID: 38.07 -> 29.50 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Face | CCSD (×10): 0.83 -> 0.35 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Cat | FID: 19.14 -> 15.67 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Cat | CCSD (×10): 0.56 -> 0.23 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Car | FID: 36.36 -> 33.50 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Car | CCSD (×10): 0.73 -> 0.37 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Horse | FID: 21.90 -> 21.18 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Horse | CCSD (×10): 1.20 -> 0.68 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Elephant | FID: 11.17 -> 10.86 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Elephant | CCSD (×10): 1.19 -> 0.82 | Helped |
| Use FreeDrag instead of DragGAN | StyleGAN2 dragging · FreeDragBench Face | time (seconds): 8.26 -> 2.74 | Helped |
| Use FreeDrag instead of DragDiffusion | Latent diffusion model dragging · DragBench | MD: 38.76 -> 33.49 | Helped |
Limitations
- Evaluation of GAN-based dragging relies on a custom FreeDragBench and a proxy CCSD metric rather than ground truth.
- Diffusion-based version requires per-image LoRA fine-tuning.
- Hyperparameters l and d must be set per model category and control edit style.
Experimental setup
Datasets: FreeDragBench DragBench
Benchmarks: FreeDragBench DragBench
Metrics: FID CCSD Mean Distance LPIPS runtime
Baselines: DragGAN DragDiffusion EditGAN
Builds on: DragGAN DragDiffusion StyleGAN2 latent diffusion models
Closest papers by meaning
- Localize, Understand, Collaborate: Semantic-Aware Dragging via Intention Reasoner (NeurIPS 2024 · similarity 0.88)
- DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing (CVPR 2024 · similarity 0.87)
- DragLoRA: Online Optimization of LoRA Adapters for Drag-based Image Editing in Diffusion Model (ICML 2025 · similarity 0.86)
- CLIPDrag: Combining Text-based and Drag-based Instructions for Image Editing (ICLR 2025 · similarity 0.86)
- 3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing (ICLR 2025 · similarity 0.86)
- DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models (ICLR 2024 · 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.