BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Hong_Kong
X-LIC-LOCATION:Asia/Hong_Kong
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:HKT
DTSTART:19911015T033000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T030657Z
LOCATION:Meeting Room S423+S424\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251218T134200
DTEND;TZID=Asia/Hong_Kong:20251218T135300
UID:siggraphasia_SIGGRAPH Asia 2025_sess155_papers_1867@linklings.com
SUMMARY:Voost: A Unified and Scalable Diffusion Transformer for Bidirectio
 nal Virtual Try-On and Try-Off
DESCRIPTION:Seungyong Lee (NXN LABS, Korea Advanced Institute of Science a
 nd Technology (KAIST)) and Jeonggi Kwak (University of British Columbia, K
 orea University)\n\nVirtual try-on aims to synthesize a realistic image of
  a person wearing a target garment, but accurately modeling garment–body c
 orrespondence remains a persistent challenge, especially under pose and ap
 pearance variation.\nIn this paper, we propose Voost—a unified and scalabl
 e framework that jointly learns virtual try-on and try-off with a single d
 iffusion transformer.\nBy modeling both tasks jointly, Voost enables each 
 garment-person pair to supervise both directions and supports flexible con
 ditioning over generation direction and garment category—enhancing garment
 –body relational reasoning without task-specific networks, auxiliary losse
 s, or additional labels.\nIn addition, we introduce two inference-time tec
 hniques: attention temperature scaling for robustness to resolution or mas
 k variation, and self-corrective sampling that leverages bidirectional con
 sistency between tasks.\nExtensive experiments demonstrate that Voost achi
 eves state-of-the-art results on both try-on and try-off benchmarks, consi
 stently outperforming strong baselines in alignment accuracy, visual fidel
 ity, and generalization.\n\nRegistration Category: Full Access, Full Acces
 s Supporter\n\nSession Chair: Ali Mahdavi-Amiri (Simon Fraser University, 
 MARZ VFX)\n\n
END:VEVENT
END:VCALENDAR