Research map / 3D vision / 3D generation
Dragin3D: Image Editing by Dragging in 3D Space
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.
Dragin3D extends drag-based image editing to 3D by representing object z-axis rotation with two tracked points rendered as circle and square Gaussian maps, injected into an SVD+ControlNet model to generate novel-view images.
Problem
Drag-based image editing is limited to 2D manipulation; existing drag methods cannot control 3D object rotation or produce novel views from simple point-drag interactions, and trajectory-based conditions are ambiguous for rotation.
Approach
Represent object rotation with a two-point line-segment trajectory projected via camera parameters into Gaussian maps of distinct shapes (circle and square) to disambiguate symmetric views, plus a CLIP-based image prompt mapping network injected into the Gaussian features, and random background augmentation of 3D object data.
Key results
- Dragin3D reaches FID 43.4, LPIPS 0.179, FVD 584.6 versus SV3D's 52.5, 0.410, 944.6 on 200 OmniObject3D test objects.
- Dragin3D achieves MD 9.38 versus DragDiffusion's 33.01 on DragBench rotation cases.
- Ablations show removing SDGS raises FVD from 584.6 to 676.8 and removing CCIM raises it to 694.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Dragin3D instead of SV3D | Multi-view image generation · OmniObject3D testset | FID: 52.5 -> 43.4 | Helped |
| Use Dragin3D instead of SV3D | Multi-view image generation · OmniObject3D testset | LPIPS: 0.410 -> 0.179 | Helped |
| Use Dragin3D instead of SV3D | Multi-view image generation · OmniObject3D testset | FVD: 944.6 -> 584.6 | Helped |
| Use Dragin3D instead of DragDiffusion | Object rotation drag editing · DragBench object rotation cases | MD: 33.01 -> 9.38 | Helped |
| Add shape distinction between the two Gaussian maps (ablation) | Dragin3D · OmniObject3D testset | FID: 44.1 -> 43.4 | Helped |
| Add shape distinction between the two Gaussian maps (ablation) | Dragin3D · OmniObject3D testset | LPIPS: 0.182 -> 0.179 | Helped |
| Add shape distinction between the two Gaussian maps (ablation) | Dragin3D · OmniObject3D testset | FVD: 676.8 -> 584.6 | Helped |
| Add image-feature injection into the Gaussian maps (ablation) | Dragin3D · OmniObject3D testset | FID: 45.2 -> 43.4 | Helped |
| Add image-feature injection into the Gaussian maps (ablation) | Dragin3D · OmniObject3D testset | LPIPS: 0.194 -> 0.179 | Helped |
| Add image-feature injection into the Gaussian maps (ablation) | Dragin3D · OmniObject3D testset | FVD: 694.1 -> 584.6 | Helped |
Limitations
- Only z-axis (left-right) rotation in one plane is supported
- Assumes z-coordinate of each pixel is fixed to 1, introducing approximation error in trajectory computation
- Trained only on synthetic 3D object dataset; real-image evaluation is qualitative
Experimental setup
Datasets: OmniObject3D
Benchmarks: OmniObject3D test subset (200 objects) DragBench
Metrics: FID LPIPS FVD MD
Baselines: DragAnything SV3D Wonder3D DragDiffusion
Builds on: ControlNet Stable Video Diffusion DragAnything DragDiffusion
Closest papers by meaning
- 3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing (ICLR 2025 · similarity 0.90)
- DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing (CVPR 2024 · similarity 0.87)
- LeviTor: 3D Trajectory Oriented Image-to-Video Synthesis (CVPR 2025 · similarity 0.87)
- FlowDrag: 3D-aware Drag-based Image Editing with Mesh-guided Deformation Vector Flow Fields (ICML 2025 · similarity 0.86)
- CTRL-D: Controllable Dynamic 3D Scene Editing with Personalized 2D Diffusion (CVPR 2025 · similarity 0.85)
- FreeDrag: Feature Dragging for Reliable Point-based Image Editing (CVPR 2024 · similarity 0.85)
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.