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 S221\, Level 2
DTSTART;TZID=Asia/Hong_Kong:20251216T110100
DTEND;TZID=Asia/Hong_Kong:20251216T111200
UID:siggraphasia_SIGGRAPH Asia 2025_sess119_papers_1730@linklings.com
SUMMARY:High-Fidelity Dynamic Portrait Animation via Direct Preference Opt
 imization and Temporal Motion Modulation
DESCRIPTION:Jiahao Cui, Baoyou Chen, Mingwang Xu, Hanlin Shang, Yuxuan Che
 n, and Qinkun Su (Fudan University); Zilong Dong (Alibaba Group); Yao Yao 
 (Nanjing University); Jingdong Wang (Baidu); and Siyu Zhu (Fudan Universit
 y, Shanghai Innovative Institute)\n\nGenerating highly dynamic and photore
 alistic portrait animations driven by audio and skeletal motion remains ch
 allenging due to the need for precise lip synchronization, natural facial 
 expressions, and high-fidelity body motion dynamics. We propose a human-pr
 eference-aligned diffusion framework that addresses these challenges throu
 gh two key innovations. First, we introduce direct preference optimization
  tailored for human-centric animation, leveraging a curated dataset of hum
 an preferences to align generated outputs with perceptual metrics for port
 rait motion-video alignment and naturalness of expression. Second, the pro
 posed temporal motion modulation resolves spatiotemporal resolution mismat
 ches by reshaping motion conditions into dimensionally aligned latent feat
 ures through temporal channel redistribution and proportional feature expa
 nsion, preserving the fidelity of high-frequency motion details in diffusi
 on-based synthesis. The proposed mechanism is complementary to existing UN
 et and DiT-based portrait diffusion approaches, and experiments demonstrat
 e obvious improvements in lip-audio synchronization, expression vividness,
  body motion coherence over baseline methods, alongside notable gains in h
 uman preference metrics. Code and datasets will be released to advance rep
 roducible research in preference-aligned portrait animation.\n\nRegistrati
 on Category: Full Access, Full Access Supporter\n\nSession Chair: Feng Xu 
 (Tsinghua University)\n\n
END:VEVENT
END:VCALENDAR