Research map / Image generation / Editing and style transfer
MetaPortrait: Identity-Preserving Talking Head Generation With Fast Personalized Adaptation
Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, HsiangTao Wu, Dong Chen, Qifeng Chen, Yong Wang, Fang Wen
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 |
|---|---|---|---|
| Use dense landmarks instead of sparse landmarks (ablation) | MetaPortrait warping network · Self-reconstruction (256 × 256) | LPIPS: 0.2485 -> 0.2262 | Improved |
| Add source identity input to refinement (ablation) | MetaPortrait refinement network · Self-reconstruction (256 × 256) | ID Loss: 0.2144 -> 0.1296 | Improved |
| Add source identity input to refinement (ablation) | MetaPortrait refinement network · Self-reconstruction (256 × 256) | LPIPS: 0.2256 -> 0.2262 | Hurt |
| Use dense landmarks instead of sparse landmarks (ablation) | MetaPortrait warping network · Self-reconstruction (256 × 256) | FID: 14.3190 -> 11.9528 | Improved |
| Use temporal 3D-convolution super-resolution rather than frame-wise GFPGAN (ablation) | MetaPortrait · Self-reconstruction (512 × 512) | E_warp: 0.0242 -> 0.0213 | Improved |
| Apply frame-wise GFPGAN super-resolution instead of bicubic upsampling (ablation) | MetaPortrait base model · Self-reconstruction (512 × 512) | E_warp: 0.0184 -> 0.0242 | Hurt |
| Use temporal super-resolution instead of bicubic upsampling (ablation) | MetaPortrait · Self-reconstruction (512 × 512) | FID: 25.5762 -> 21.4974 | Improved |
| Use temporal super-resolution instead of frame-wise GFPGAN (ablation) | MetaPortrait · Self-reconstruction (512 × 512) | LPIPS: 0.2178 -> 0.2079 | Improved |
| Initialize personalization with Reptile meta-learning | MetaPortrait personalization · HDTF test set | LPIPS: 0.5 epoch; 3× speedup against our base model, and 4× against FOMM | Improved |
| Use MetaPortrait instead of StyleHEAT | Talking-head generation · Self Reconstruction (512 × 512) | FID: 44.5207 -> 21.4974 | Improved |
| Use MetaPortrait instead of StyleHEAT | Talking-head generation · Cross Reenactment (512 × 512) | ID Loss: 0.4720 -> 0.1952 | Improved |
Closest papers by meaning
- Learned Spatial Representations for Few-Shot Talking-Head Synthesis (ICCV 2021 · similarity 0.88)
- High-Fidelity and Freely Controllable Talking Head Video Generation (CVPR 2023 · similarity 0.88)
- MimicTalk: Mimicking a personalized and expressive 3D talking face in minutes (NeurIPS 2024 · similarity 0.87)
- InsTaG: Learning Personalized 3D Talking Head from Few-Second Video (CVPR 2025 · similarity 0.87)
- GoHD: Gaze-oriented and Highly Disentangled Portrait Animation with Rhythmic Poses and Realistic Expressions (AAAI 2025 · similarity 0.86)
- AE-NeRF: Audio Enhanced Neural Radiance Field for Few Shot Talking Head Synthesis (AAAI 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.
