Research map / 3D vision / Face and shape
3D FaceShell: Attribute Transfer in 3D Face Avatars as a VLM Defense Mechanism
Weston Bondurant, Srijan Das, Hieu Le, Stephanie Schuckers
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add data augmentation (ablation) | 3D FaceShell · Celebrity 3D face avatars | injection rate (%) VideoLLaMA3: 17.6 -> 45.0 | Improved |
| add chroma loss (ablation) | 3D FaceShell · Celebrity 3D face avatars | identity similarity: 0.5695 -> 0.7507 | Improved |
| add chroma loss (ablation) | 3D FaceShell · Celebrity 3D face avatars | LPIPS: 0.1911 -> 0.1513 | Improved |
| add landmark consistency loss (ablation) | 3D FaceShell · Celebrity 3D face avatars | identity similarity: 0.6184 -> 0.7507 | Improved |
| add dynamic opacity clamping (ablation) | 3D FaceShell · Celebrity 3D face avatars | LPIPS: 0.1894 -> 0.1513 | Improved |
| add dynamic opacity clamping (ablation) | 3D FaceShell · Celebrity 3D face avatars | injection rate (%) LLaVA-NeXT: 31.5 -> 29.2 | Hurt |
| use SigLIP and CLIP ensemble instead of CLIP alone (ablation) | 3D FaceShell · Celebrity 3D face avatars | injection rate (%) VideoLLaMA3: 21.6 -> 45.0 | Improved |
| use SigLIP and CLIP ensemble instead of SigLIP alone (ablation) | 3D FaceShell · Celebrity 3D face avatars | injection rate (%) BLIP-2: 20.4 -> 25.7 | Improved |
| multi-view training instead of front view only (ablation) | 3D FaceShell · Celebrity 3D face avatars (validation view) | injection rate (%) VideoLLaMA3: 23.4 -> 36.4 | Improved |
| shift all attributes at once instead of single attribute (ablation) | 3D FaceShell · VideoLLaMA3 attributes | hair color transfer rate: +50.7% | Improved |
| shift all attributes at once instead of single attribute (ablation) | 3D FaceShell · VideoLLaMA3 attributes | expression transfer rate: +96.7% | Improved |
| increase shell Gaussians from 50,000 to 75,000 (ablation) | 3D FaceShell · Celebrity 3D face avatars | injection rate (%) VideoLLaMA3: 48.7 -> 52.1 | Improved |
Closest papers by meaning
- Instance-Level Facial Attributes Transfer with Geometry-Aware Flow (AAAI 2019 · similarity 0.87)
- Pixel Codec Avatars (CVPR 2021 · similarity 0.86)
- Bringing Your Portrait to 3D Presence (CVPR 2026 · similarity 0.86)
- ID-to-3D: Expressive ID-guided 3D Heads via Score Distillation Sampling (NeurIPS 2024 · similarity 0.85)
- FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs (CVPR 2025 · similarity 0.85)
- Arc2Avatar: Generating Expressive 3D Avatars from a Single Image via ID Guidance (CVPR 2025 · 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.
