Research map / 3D vision / Face and shape
HAVE-FUN: Human Avatar Reconstruction from Few-Shot Unconstrained 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.
Reconstructs animatable human body and hand avatars from 2-8 unconstrained casual images by combining a skinned drivable DMTet tetrahedral representation with reference reconstruction losses and Zero123-based SDS guidance.
Problem
Human avatar reconstruction usually needs expensive multi-view or video capture; this paper targets animatable avatars from very few (2-8) unconstrained photos with varied poses.
Approach
Integrates SMPL-X/MANO skinning weights and expression blendshapes into DMTet so the tetrahedral mesh becomes drivable and can be initialized from a canonical SMPL-X template, and uses a two-phase optimization: few-shot reference reconstruction losses plus few-shot guidance via Zero123 SDS for unseen regions.
Key results
- On FS-XHumans, HaveFun reaches PSNR 24.0/25.6/26.8 for 2/4/8-shot versus 21.0 for TeCH (1-shot) and 19.9/20.7 for SelfRecon (8-/100-shot).
- On FS-DART hand reconstruction, HaveFun achieves PSNR 26.6/26.7/26.3 for 2/4/8-shot, compared with 20.8-21.7 for SelfRecon.
- Ablations show SDS weight is task-dependent (lambda_sds = 0.05 for 2-shot body-hand setting, 0.01 for 4- and 8-shot body) and removing SDS leaves unseen textures unreasonable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HaveFun instead of SelfRecon or TeCH for few-shot reconstruction | FS-XHumans body · FS-XHumans | PSNR: 21.0 -> 24.0/25.6/26.8 | Helped |
| Use HaveFun instead of SelfRecon for few-shot hand reconstruction | FS-DART · FS-DART | PSNR: 20.8/21.3/21.7 -> 26.6/26.7/26.3 | Helped |
| Add SDS guidance with λsds = 0.01 (ablation) | HaveFun body reconstruction · 4-shot FS-XHumans | PSNR: 24.54 -> 25.64 | Helped |
| Increase SDS weight to λsds = 0.05 (ablation) | HaveFun body reconstruction · 4-shot FS-XHumans | PSNR: 25.64 -> 25.10 | Hurt |
| Increase SDS weight to λsds = 1 (ablation) | HaveFun body reconstruction · 4-shot FS-XHumans | PSNR: 24.54 -> 24.56 | Hurt |
| Add SDS guidance with λsds = 0.05 (ablation) | HaveFun hand reconstruction · 2-shot FS-DART | PSNR: 25.62 -> 26.61 | Helped |
| Increase SDS weight to λsds = 1 (ablation) | HaveFun hand reconstruction · 2-shot FS-DART | LPIPS: 0.0577 -> 0.0736 | Hurt |
| Increase training images from 2-shot to 8-shot (ablation) | HaveFun body reconstruction · FS-XHumans | PSNR: 24.0/25.6/26.8 | Helped |
| Increase training images from 2-shot to 8-shot (ablation) | HaveFun hand reconstruction · FS-DART | PSNR: 26.6/26.7/26.3 | Mixed |
Surprising findings
- For hand reconstruction, some 2-shot results are better than 8-shot results, attributed to palm/back dominating hand appearance.
Limitations
- Expression blendshapes from SMPL-X are imperfect, hurting portrait reconstruction with diverse expressions.
- Evaluation relies on synthetic benchmarks (FS-XHumans and FS-DART) with extremely articulated self-occlusion excluded; real-world results are only qualitative.
Experimental setup
Datasets: XHumans DART FS-XHumans FS-DART
Benchmarks: FS-XHumans FS-DART
Metrics: PSNR SSIM LPIPS
Baselines: SelfRecon TeCH
Builds on: DMTet Zero123 SMPL-X DreamFusion SDS Get3D
Closest papers by meaning
- NoPo-Avatar: Generalizable and Animatable Avatars from Sparse Inputs without Human Poses (NeurIPS 2025 · similarity 0.86)
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.84)
- FRESA: Feedforward Reconstruction of Personalized Skinned Avatars from Few Images (CVPR 2025 · similarity 0.83)
- GeoAvatar: Geometrically-Consistent Multi-Person Avatar Reconstruction from Sparse Multi-View Videos (CVPR 2025 · similarity 0.83)
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.82)
- GAF: Gaussian Avatar Reconstruction from Monocular Videos via Multi-view Diffusion (CVPR 2025 · similarity 0.82)
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.