Answers are grounded in this paper and cite their sources.
Reconstructing 3D human avatars from monocular video typically faces the challenge of appearance collapse caused by incomplete observation. Existing methods commonly utilise pretrained image diffusion models as generative priors, distilling structural and texture knowledge to supervise 3D generation. However, these priors often struggle to recover subject‐specific details, and injecting them comes at the cost of distorting observed regions. To address these issues, this paper presents a dual‐branch framework, termed CompAvatar, which can generate and fuse semantically consistent views in missing observations for improved 3D avatar integrity. Specifically, an observation modelling branch reconstructs coarse 3D Gaussians to guide the view augmentation branch. To address the challenge of refining appearance without distorting observed regions within view augmentation, CompAvatar introduces three modules. First, to ensure semantic consistency, the component semantic learning module extracts fine‐grained components from video to fine‐tune a personalised view generator, capturing specific‐subject appearance. Subsequently, to restore details while protecting observed regions, the localised view refinement module employs component‐level anomaly detection to locate defect masks for view generation. Finally, the spatial consistency distillation strategy leverages 3D spatial consistency to progressively feed generated views back into the 3D representation. Experiments show CompAvatar outperforms existing methods in restoring missing details from partial‐view monocular videos.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: