Culture & Life Yonhap News Culture · 2026-10-27

South Korean Researchers Unveil 'FlexiAvatar' AI to Craft Ultra-Crisp 3D Avatars from Any Camera Angle

Key Points
  • Researchers at UNIST and the University of Aberdeen have developed 'FlexiAvatar', an AI system that generates precise 3D avatars across face-only, upper-body, and full-body footage within a single pip
  • By calculating the visibility of body regions and focusing computational power only on observed features, the technology prevents the blurring and distortion common in older reconstruction models.
  • Presented at the European Conference on Computer Vision (ECCV), the tech holds significant potential for virtual reality, teleconferencing, gaming, and digital entertainment.
Summary South Korean researchers have unveiled a breakthrough artificial intelligence framework that can reconstruct lifelike 3D avatars regardless of whether a camera captures only a user's face, upper torso, or entire body. Developed by a joint team led by Professor Baek Seung-ryeol at the Ulsan National Institute of Science and Technology (UNIST) Graduate School of Artificial Intelligence alongside researchers from the University of Aberdeen, the new technology—dubbed "FlexiAvatar"—promises to eliminate common visual artifacts and computationally intensive errors that have long troubled virtual avatar creation. Traditionally, 3D body reconstruction models have struggled with varying camera framing. Conventional full-body AI models attempt to calculate and flesh out unseen body parts, such as legs hidden beneath a desk during a webcam call. This unnecessary estimation often introduces distortion, blurring or warping well-visible areas like the face or chest. While specialized models focused strictly on headshots or upper bodies exist, digital creators previously had to switch between completely different tools whenever camera angles or framing changed. FlexiAvatar solves this friction by analyzing how frequently each small 3D component of an avatar appears across video frames. The AI allocates its computational power toward refining the shape and coloration of clearly observed regions while minimizing wasteful calculations for uncaptured limbs. To guarantee expressiveness, it incorporates dedicated refinement pipelines specifically for facial expressions and subtle hand gestures. For angles that remain unseen—such as the back of the head or shoulders—the system leverages generative AI to construct plausible complementary views, carefully balancing them so actual captured video data always takes precedence. In standard academic benchmarks, FlexiAvatar outperformed existing state-of-the-art models. Tested on the NeuMan benchmark, it achieved a Peak Signal-to-Noise Ratio (PSNR) of 35.77 dB, surpassing comparative technology ExAvatar (34.80 dB), with higher numbers denoting fidelity closer to the original source footage. The technology is poised to accelerate virtual human production across South Korea's vibrant digital content ecosystem, spanning virtual reality platforms, remote teleconferencing, video games, and film visual effects.
Sponsored · Ad