Researchers have developed AHOY, a novel method for creating complete, animatable 3D Gaussian avatars from monocular YouTube videos, even when subjects are heavily occluded. The system addresses challenges like unobserved body regions and lack of multi-view supervision by employing a hallucination-as-supervision pipeline using identity-finetuned diffusion models. A two-stage architecture and a map-pose/LBS-pose decoupling strategy further enhance reconstruction quality, with a head/body split supervision preserving facial identity. The resulting avatars are demonstrated to be robust enough for animation and integration into 3D scenes. AI
IMPACT Enables creation of detailed 3D avatars from everyday video, potentially impacting virtual reality, gaming, and digital content creation.
RANK_REASON The cluster describes a new research paper detailing a novel method for 3D avatar reconstruction. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
