Research map / Continual learning / Online and replay-based
MaintaAvatar: A Maintainable Avatar Based on Neural Radiance Fields by Continual Learning
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 MaintaAvatar, a continual-learning NeRF framework for human avatars that learns new appearances and poses from few images while avoiding forgetting, using a Global-Local Joint Storage Module and a Pose Distillation Module.
Problem
Updating a NeRF-based human avatar with new appearances and poses from few images without catastrophic forgetting of past appearances, including color bleeding and incorrect rendered poses for past tasks.
Approach
Combines generative replay (frozen previous-task network supervises current training) with a joint global-local embedding storage (global geometry/color embeddings plus per-appearance tri-plane local embeddings generated from learned condition embeddings) to separate appearances, and distills the pose-correction MLP outputs from the previous model to retain past poses.
Key results
- On ZJU-MoCap novel view, Ours reaches 29.495 PSNR and 0.9663 SSIM versus 28.605 PSNR for PersonNeRF_CL and 25.810 for MEIL-NeRF.
- On THuman2.0 (subject 207), Ours achieves 23.998 PSNR and 0.9664 SSIM versus 22.951 PSNR for PersonNeRF_CL.
- Ablation on ZJU-MoCap novel view: removing the pose distillation module drops PSNR from 29.50 to 29.21 and removing the global-local module drops it to 28.89.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the Pose Distillation Module (ablation) | MaintaAvatar · ZJU-MoCap novel view | PSNR: 29.21 -> 29.50 | Helped |
| Add the Pose Distillation Module (ablation) | MaintaAvatar · ZJU-MoCap novel pose | PSNR: 29.59 -> 29.63 | Helped |
| Add the Global-Local Joint Storage Module (ablation) | MaintaAvatar · ZJU-MoCap novel view | PSNR: 28.89 -> 29.50 | Helped |
| Add the Global-Local Joint Storage Module (ablation) | MaintaAvatar · ZJU-MoCap novel pose | PSNR: 29.15 -> 29.63 | Helped |
Limitations
- Performance drops with significant clothing shape changes.
- Struggles with pose generalization in complex poses due to limited few-shot exposure.
Experimental setup
Datasets: ZJU-MoCap THuman2.0
Benchmarks: ZJU-MoCap THuman2.0
Metrics: PSNR SSIM LPIPS
Baselines: Joint CLNeRF MEIL-NeRF PersonNeRF PersonNeRF_CL
Builds on: NeRF PersonNeRF MEIL-NeRF CLNeRF
Closest papers by meaning
- Artist-Friendly Relightable and Animatable Neural Heads (CVPR 2024 · similarity 0.86)
- RAM-Avatar: Real-time Photo-Realistic Avatar from Monocular Videos with Full-body Control (CVPR 2024 · similarity 0.86)
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.85)
- High-Fidelity 3D Head Avatars Reconstruction through Spatially-Varying Expression Conditioned Neural Radiance Field (AAAI 2024 · similarity 0.85)
- RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars (CVPR 2025 · similarity 0.84)
- Relightable and Animatable Neural Avatars from Videos (AAAI 2024 · similarity 0.84)
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.