Research map / 3D vision / Face and shape
3D-Aware Face Editing via Warping-Guided Latent Direction Learning
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 FaceEdit3D, which lets users drag 3D facial landmarks, warps EG3D tri-planes, and uses a warp-aware encoder plus hierarchical directional latent editing to produce identity-preserving 3D face edits in real time.
Problem
Interactive 3D-aware face editing that is intuitive (point dragging), generalizes without per-edit retraining, and runs efficiently, which prior 2D drag methods (DragGAN) and 3D fine-tuning or conditional methods do not jointly satisfy.
Approach
Warp the EG3D tri-plane features with thin-plate spline interpolation driven by user-dragged 3D landmarks, then train a warp-aware encoder to map the distorted renderings back to a standardized W+ latent space, and learn editing directions as differences of encoder latents before and after warping, selected per layer group.
Key results
- Editing inference time of 0.624 s for the 3D method versus 5.231 s for DragGAN (2D), with ID consistency 0.712 vs 0.579 for DragGAN.
- MSE inside/outside ratio of 4.909 for the proposed method versus 1.865 for IDE-3D, indicating better disentangled editing.
- When extended to the same 2D generator, the method reportedly performs better than DragGAN, though the 3D version falls behind DragGAN in editability.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FaceEdit3D instead of IDE-3D | 3D face editing · 3D editing | MSEi / MSEo: 1.865 -> 4.909 | Helped |
| Use FaceEdit3D instead of IDE-3D | 3D face editing · 3D editing | ID Consistency: 0.649 -> 0.712 | Helped |
| Use FaceEdit3D instead of DragGAN | 2D face editing · 2D editing | Inference Time (s): 5.231 -> 0.356 | Helped |
| Add tri-plane warping as data augmentation when training the encoder (ablation) | FaceEdit3D warp-aware encoder · facial point manipulation | manipulation perception | Helped |
| Add the discriminator loss Ld during encoder training (ablation) | FaceEdit3D warp-aware encoder · face editing | identity preservation | Helped |
| Combine warp-assisted data augmentation and discriminator loss in encoder training (ablation) | FaceEdit3D warp-aware encoder · face editing | editing quality | Helped |
| Add hierarchical directional latent editing (ablation) | FaceEdit3D · face editing | identity preservation and detail | Helped |
| Add hierarchical directional latent editing (ablation) | FaceEdit3D · face editing | attribute disentanglement | Helped |
Surprising findings
- Direct tri-plane warping is fastest but causes facial distortion, which motivates the warp-aware encoder.
- The 3D method trails DragGAN in editability, which the authors attribute to lower editability of 3D GANs compared with 2D GANs.
Limitations
- Warping the 3D representation cannot handle texture editing.
- Semantic edits such as adding glasses are not supported.
- The 3D generator's editability is weaker than 2D GANs, so results lag DragGAN in editing strength.
Experimental setup
Datasets: FFHQ AFHQ Cats Stanford Cars
Benchmarks: FFHQ
Metrics: inference time MSE inside edited region MSE outside edited region ID consistency (ArcFace)
Baselines: DragGAN IDE-3D StyleGAN-NADA direct tri-plane warp
Builds on: EG3D DragGAN GOAE StyleCLIP thin-plate spline
Closest papers by meaning
- StrokeFaceNeRF: Stroke-based Facial Appearance Editing in Neural Radiance Field (CVPR 2024 · similarity 0.86)
- Fast Text-to-3D-Aware Face Generation and Manipulation via Direct Cross-modal Mapping and Geometric Regularization (ICML 2024 · similarity 0.86)
- Reference-Based 3D-Aware Image Editing with Triplanes (CVPR 2025 · similarity 0.85)
- FFaceNeRF: Few-shot Face Editing in Neural Radiance Fields (CVPR 2025 · similarity 0.85)
- InstructPix2NeRF: Instructed 3D Portrait Editing from a Single Image (ICLR 2024 · similarity 0.85)
- VOODOO 3D: Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment (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.