BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Tokyo
X-LIC-LOCATION:Asia/Tokyo
BEGIN:STANDARD
TZOFFSETFROM:+0900
TZOFFSETTO:+0900
TZNAME:JST
DTSTART:18871231T000000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20250110T023313Z
LOCATION:Hall B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241206T110800
DTEND;TZID=Asia/Tokyo:20241206T111900
UID:siggraphasia_SIGGRAPH Asia 2024_sess143_papers_1054@linklings.com
SUMMARY:GGHead: Fast and Generalizable 3D Gaussian Heads
DESCRIPTION:Technical Papers\n\nTobias Kirschstein, Simon Giebenhain, and 
 Jiapeng Tang (Technical University of Munich); Markos Georgopoulos (Indepe
 ndent); and Matthias Nießner (Technical University of Munich)\n\nLearning 
 3D head priors from large 2D image collections is an important step toward
 s high-quality 3D-aware human modeling. \nA core requirement is an efficie
 nt architecture that scales well to large-scale datasets and large image r
 esolutions. \nUnfortunately, existing 3D GANs struggle to scale to generat
 e samples at high resolutions due to their relatively slow train and rende
 r speeds, and typically have to rely on 2D superresolution networks at the
  expense of global 3D consistency. \nTo address these challenges, we propo
 se Generative Gaussian Heads (GGHead), which adopts the recent 3D Gaussian
  Splatting representation within a 3D GAN framework. \nTo generate a 3D re
 presentation, we employ a powerful 2D CNN generator to predict Gaussian at
 tributes in the UV space of a template head mesh. \nThis way, GGHead explo
 its the regularity of the template's UV layout, substantially facilitating
  the challenging task of predicting an unstructured set of 3D Gaussians. \
 nWe further improve the geometric fidelity of the generated 3D representat
 ions with a novel total variation loss on rendered UV coordinates. \nIntui
 tively, this regularization encourages that neighboring rendered pixels sh
 ould stem from neighboring Gaussians in the template’s UV space. \nTaken t
 ogether, our pipeline can efficiently generate 3D heads trained only from 
 single-view 2D image observations. \nOur proposed framework matches the qu
 ality of existing 3D head GANs on FFHQ while being both substantially fast
 er and fully 3D consistent. \nAs a result, we demonstrate real-time genera
 tion and rendering of high-quality 3D-consistent heads at 1024x1024 resolu
 tion for the first time.\n\nRegistration Category: Full Access, Full Acces
 s Supporter\n\nLanguage Format: English Language\n\nSession Chair: Iain Ma
 tthews (Epic Games, Carnegie Mellon University)
URL:https://asia.siggraph.org/2024/program/?id=papers_1054&sess=sess143
END:VEVENT
END:VCALENDAR