Research map / 3D vision / Depth and stereo
Face Anything: 4D Face Reconstruction from Any Image Sequence
Umut Kocasarı, Simon Giebenhain, Richard Shaw, Matthias Niessner
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 |
|---|---|---|---|
| predict canonical maps instead of motion to canonical space (ablation) | Face Anything · NeRSemble | EPE(2D): 6.210 -> 3.271 | Improved |
| sample both cameras and timestamps instead of fixed timestamp (static training) (ablation) | Face Anything · NeRSemble | EPE(2D): 3.864 -> 3.271 | Improved |
| sample both cameras and timestamps instead of fixed camera (monocular training) (ablation) | Face Anything · NeRSemble 16-view depth | AbsRel (x10): 0.064 -> 0.015 | Improved |
| sample both cameras and timestamps instead of fixed camera (monocular training) (ablation) | Face Anything · NeRSemble | EPE(2D): 3.031 (monocular training) vs 3.271 (full) | Hurt |
| add DAViD facial pretraining to DA3 (ablation) | DA3 · NeRSemble | AbsRel monocular (x10): 0.085 -> 0.061 | Improved |
| add multi-view finetuning with canonical head on top of DAViD pretraining (ablation) | DAViD Pretrained · NeRSemble 16 views | AbsRel (x10): 0.033 -> 0.015 | Improved |
| use multiple input images instead of a single image | Face Anything · NeRSemble | depth error | Improved |
| canonical map prediction vs V-DPM motion-based method | Face Anything · Efficiency | runtime (s): 160 -> 5 | Improved |
| Face Anything vs P3DMM on hair regions | Face Anything · NeRSemble 2D correspondence margin 2 | EPE(2D): 5.550 -> 1.838 | Improved |
| Face Anything vs V-DPM | Face Anything · VFHQ margin 5 | CCE mean: 0.797 -> 0.398 | Improved |
| Face Anything vs Pi3 and Sapiens-1B on Ava-256 video depth | Face Anything · Ava-256 monocular depth | RMSE (x10): 0.081 -> 0.067 (best baseline Sapiens-2B) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- V-DPM — V-DPM: 4D Video Reconstruction with Dynamic Point Maps (CVPR 2026) · better on 7 of 7 shared results
- Sapiens-2B — Sapiens: Foundation for Human Vision Models (ECCV 2024) · better on 4 of 4 shared results
- DAViD — DAViD: Modeling Dynamic Affordance of 3D Objects Using Pre-trained Video Diffusion Models (ICCV 2025) · better on 1 of 1 shared result
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
- Large Pose 3D Face Reconstruction From a Single Image via Direct Volumetric CNN Regression (ICCV 2017 · similarity 0.88)
- Towards Metrical Reconstruction of Human Faces (ECCV 2022 · similarity 0.88)
- AVFace: Towards Detailed Audio-Visual 4D Face Reconstruction (CVPR 2023 · similarity 0.88)
- AU-Aware 3D Face Reconstruction through Personalized AU-Specific Blendshape Learning (ECCV 2022 · similarity 0.88)
- 3D Face Tracking from 2D Video through Iterative Dense UV to Image Flow (CVPR 2024 · similarity 0.87)
- PSHuman: Photorealistic Single-image 3D Human Reconstruction using Cross-Scale Multiview Diffusion and Explicit Remeshing (CVPR 2025 · similarity 0.87)
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.
