Research map / 3D vision / 3D generation
Portrait4D-v2: Pseudo Multi-View Data Creates Better 4D Head Synthesizer
Yu Deng, Duomin Wang, Baoyuan Wang
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 |
|---|---|---|---|
| initialize 4D synthesizer from pretrained 3D synthesizer instead of random (ablation) | Portrait4D-v2 · VFHQ cross-ID reenactment | APD: 1.887 -> 0.796 | Improved |
| add pseudo multi-view driving frame I_d(theta_p) for motion embedding (ablation) | Portrait4D-v2 · VFHQ cross-ID reenactment | identity consistency (Cons.): 0.813 -> 0.913 | Improved |
| add pseudo multi-view supervision I_d(theta_q) (ablation) | Portrait4D-v2 · VFHQ cross-ID reenactment | APD: 0.796 -> 0.688 | Improved |
| use implicit 2D motion embedding instead of 3DMM parameters (ablation) | Portrait4D-v2 · VFHQ cross-ID reenactment | AED_f: 0.202 -> 0.186 | Improved |
| use own 3D synthesizer instead of Portrait4D for multi-view data (ablation) | Portrait4D-v2 · VFHQ cross-ID reenactment | AED_f: 0.188 -> 0.186 | Improved |
| increase training videos from 1K to 10K (ablation) | Portrait4D-v2 · VFHQ self reenactment | ID: 0.580 -> 0.873 | Improved |
| train with fewer videos (1K-2K) on cross-ID AED and APD (ablation) | Portrait4D-v2 · VFHQ cross-ID reenactment | APD: 0.700 (1K) vs 0.943 (10K) | Mixed |
| raise motion-layer learning rate by 2.5x and train on 6M images (ablation) | Portrait4D-v2 · VFHQ cross-ID reenactment | APD: 0.943 -> 0.858 | Improved |
| pseudo multi-view learning versus prior one-shot head avatar methods | Portrait4D-v2 · VFHQ self reenactment | FID: 43.0 -> 26.3 vs Portrait4D | Improved |
| Portrait4D-v2 versus PIRenderer on cross-ID AED metric | Portrait4D-v2 · VFHQ cross-ID reenactment | AED: 0.161 vs 0.146 (PIRenderer) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Portrait4D — Portrait4D: Learning One-Shot 4D Head Avatar Synthesis using Synthetic Data (CVPR 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Portrait4D: Learning One-Shot 4D Head Avatar Synthesis using Synthetic Data (CVPR 2024 · similarity 0.94)
- CAP4D: Creating Animatable 4D Portrait Avatars with Morphable Multi-View Diffusion Models (CVPR 2025 · similarity 0.89)
- Synthetic Prior for Few-Shot Drivable Head Avatar Inversion (CVPR 2025 · similarity 0.87)
- IM-Portrait: Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC (CVPR 2025 · similarity 0.87)
- FA-LAM: Focus-Aware Large Avatar Model for One-Shot 4D Animatable Gaussian Head (NeurIPS 2026 · similarity 0.86)
- Coherent 3D Portrait Video Reconstruction via Triplane Fusion (CVPR 2025 · 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.
