Research map / 3D vision / Human pose estimation
NoPo-Avatar: Generalizable and Animatable Avatars from Sparse Inputs without Human Poses
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.
Proposes NoPo-Avatar, a feed-forward model that reconstructs a canonical T-pose animatable 3D human avatar from sparse images without any camera or human pose inputs, using a template branch plus pixel-aligned image branches.
Problem
Generalizable animatable human avatar reconstruction from sparse or single images normally requires accurate camera and SMPL-X poses at test time, and rendering quality degrades sharply when these poses are noisy (e.g., predicted by off-the-shelf estimators).
Approach
Eliminate pose inputs from reconstruction entirely: an encoder-decoder with a learnable SMPL-X T-pose template branch (for inpainting unseen regions) and per-view pixel-aligned image branches (for observed details), all predicted in canonical T-pose space and animated via LBS, trained end-to-end with photometric, chamfer, projection and LBS-weight losses.
Key results
- On THuman2.0 with 3 input views and predicted poses for baselines, the method reaches 22.49 PSNR / 105.45 LPIPS* / 42.19 FID versus LIFe-GoM's 19.70 / 146.19 / 63.34 without test-time optimization.
- On THuman2.0 with ground-truth poses for baselines, the method gets 22.49 PSNR, slightly below LIFe-GoM's 24.65, but better LPIPS* (105.45 vs 110.82) and FID (42.19 vs 51.27).
- Single-image HuGe100K: 23.15 PSNR / 90.63 LPIPS* / 15.56 FID versus IDOL's 20.89 / 111.68 / 16.91, with 321.58ms reconstruction time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| reconstruct from images without input pose priors instead of using predicted input poses | NoPo-Avatar vs LIFe-GoM · THuman2.0 sparse-input novel view synthesis, predicted poses, no test-time optimization | LPIPS*: 146.19 -> 105.45 | Helped |
| reconstruct from images without input pose priors instead of using predicted input poses | NoPo-Avatar vs LIFe-GoM · THuman2.0 sparse-input novel view synthesis, predicted poses, no test-time optimization | PSNR: 19.70 -> 22.49 | Helped |
| reconstruct from images without input pose priors instead of using ground-truth input poses | NoPo-Avatar vs LIFe-GoM · THuman2.0 sparse-input novel view synthesis, ground-truth input poses | LPIPS*: 110.82 -> 105.45 | Helped |
| use NoPo-Avatar rather than IDOL | single-image novel view synthesis · HuGe100K | PSNR: 20.89 -> 23.15 | Helped |
| use NoPo-Avatar rather than IDOL | single-image novel view synthesis · THuman2.0 | LPIPS*: 66.62 -> 49.69 | Helped |
| add the template branch to image branches (ablation) | NoPo-Avatar · THuman2.0 ablation, 1 input | FID: 66.53 -> 57.78 | Helped |
| add image branches to the template branch (ablation) | NoPo-Avatar · THuman2.0 ablation, 3 inputs | FID: 47.60 -> 42.18 | Helped |
| remove projection loss L_proj (ablation) | NoPo-Avatar · THuman2.0 auxiliary-loss ablation | FID: 42.18 -> 48.98 | Hurt |
| remove LBS loss L_lbs (ablation) | NoPo-Avatar · novel pose synthesis | novel pose synthesis quality | Hurt |
| remove LBS loss L_lbs (ablation) | NoPo-Avatar · novel view synthesis | novel view synthesis quality | No clear effect |
| remove Chamfer loss L_chamfer (ablation) | NoPo-Avatar · final model | final model performance | No clear effect |
| increase training data from THuman2.0 to THuman2.1 | NoPo-Avatar · XHuman cross-domain novel view synthesis | PSNR: +0.6 | Helped |
Surprising findings
- Training LIFe-GoM with noisy poses does not improve its robustness to noisy poses at test time; training on ground-truth poses often works best.
- Scaling training data from THuman2.0 to THuman2.1 improves the pose-free method on cross-domain XHuman, while LIFe-GoM does not scale.
Limitations
- PSNR is slightly lower than GT-pose baselines such as LIFe-GoM in the lab setting (22.49 vs 24.65).
- Training data (THuman2.0/2.1) contains only Asian subjects, so generalization to other ethnicities may be limited.
- HuGe100K's diffusion-synthesized multiview inconsistency limits gains from scaling in LPIPS* and FID.
Experimental setup
Datasets: THuman2.0 THuman2.1 HuGe100K XHuman
Benchmarks: THuman2.0 XHuman HuGe100K
Metrics: PSNR LPIPS* FID
Baselines: GHG LIFe-GoM GoMAvatar 3DGS-Avatar iHuman NHP NIA IDOL LHM
Builds on: NoPoSplat DUSt3R LIFe-GoM IDOL SMPL-X
Closest papers by meaning
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.88)
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian Reconstruction (CVPR 2025 · similarity 0.87)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.87)
- 3D²-Actor: Learning Pose-Conditioned 3D-Aware Denoiser for Realistic Gaussian Avatar Modeling (AAAI 2025 · similarity 0.87)
- Relightable and Animatable Neural Avatar from Sparse-View Video (CVPR 2024 · similarity 0.86)
- FRESA: Feedforward Reconstruction of Personalized Skinned Avatars from Few Images (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.