Research map / 3D vision / Face and shape
FRESA: Feedforward Reconstruction of Personalized Skinned Avatars from Few Images
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.
Feed-forward method that reconstructs personalized skinned 3D human avatars (canonical shape, personalized skinning weights, pose-dependent deformation) from a few front/back posed images via a universal clothed-human model trained on a 1100-subject dome dataset.
Problem
Reconstructing animatable personalized 3D clothed human avatars from only a few casually captured images currently requires hours of per-subject optimization, and fixed/nearest-vertex skinning causes deformation artifacts.
Approach
Learn a universal prior over many clothed humans that jointly infers personalized canonical geometry (DMTet on a tetrahedral grid), personalized skinning weights, and pose-dependent displacements, using 3D canonicalization (unposed normal/segmentation images) and multi-frame averaged bi-plane features, trained in canonical-space then posed-space stages.
Key results
- On Dome Data, the full model reaches P2S 0.37 cm and CD 0.43 cm versus 0.98/1.12 cm for Vid2Avatar and 1.47/1.63 cm for PuzzleAvatar.
- Inference takes about 18 s per avatar, compared with 8 h for Vid2Avatar and 3 h for PuzzleAvatar.
- On RenderPeople, the full model reaches P2S 0.30 cm, versus 2.44 cm for ARCH++ and 1.79 cm for PuzzleAvatar.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pose-dependent deformation (ablation) | FRESA · Dome Data | Normal: 0.030 -> 0.026 | Helped |
| add pose-dependent deformation (ablation) | FRESA · Dome Data | P2S: 0.43 -> 0.37 | Helped |
| add pose-dependent deformation (ablation) | FRESA · Dome Data | CD: 0.49 -> 0.43 | Helped |
| add pose-dependent deformation (ablation) | FRESA · RenderPeople | Normal: 0.026 -> 0.024 | Helped |
| add pose-dependent deformation (ablation) | FRESA · RenderPeople | P2S: 0.33 -> 0.30 | Helped |
| add pose-dependent deformation (ablation) | FRESA · RenderPeople | CD: 0.38 -> 0.34 | Helped |
| canonicalize inputs instead of sampling features from posed inputs via forward warping (ablation) | FRESA · Canonical geometry reconstruction | geometric fidelity | Helped |
| aggregate multiple reference frames instead of using one frame (ablation) | FRESA · Canonical geometry reconstruction | plausibility | Helped |
| predict personalized skinning weights instead of using nearest-template weights (ablation) | FRESA · Avatar animation | deformation artifacts | Helped |
| train with multiple input and target frames (N=5) instead of the same input and target frame (N=1) (ablation) | FRESA skinning weights · Avatar animation | skinning-weight robustness | Helped |
| add pose-dependent deformation (ablation) | FRESA · Avatar animation | animation realism | Helped |
Limitations
- Geometric fidelity bounded by tetrahedral grid resolution, missing tiny accessories.
- Only pose-dependent deformations modeled; body-cloth dynamics, long hair, and loose garments not captured.
Experimental setup
Datasets: Dome Data (1100 subjects, in-house) RenderPeople (130 synthetic subjects, test)
Benchmarks: Dome Data (35 unseen subjects) RenderPeople
Metrics: Normal error P2S Chamfer distance inference time
Baselines: ARCH++ Vid2Avatar PuzzleAvatar
Builds on: ARCH++ PuzzleAvatar DMTet LBS Vid2Avatar
Closest papers by meaning
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.90)
- DLCA-Recon: Dynamic Loose Clothing Avatar Reconstruction from Monocular Videos (AAAI 2024 · similarity 0.88)
- GeoAvatar: Geometrically-Consistent Multi-Person Avatar Reconstruction from Sparse Multi-View Videos (CVPR 2025 · similarity 0.87)
- MPMAvatar: Learning 3D Gaussian Avatars with Accurate and Robust Physics-Based Dynamics (NeurIPS 2025 · similarity 0.87)
- CtrlAvatar: Controllable Avatars Generation via Disentangled Invertible Networks (AAAI 2025 · similarity 0.87)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · 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.