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:20260817T171531Z
LOCATION:Hall B7 (1)\, B Block\, Level 7
DTSTART;TZID=Asia/Tokyo:20241203T130000
DTEND;TZID=Asia/Tokyo:20241203T131100
UID:siggraphasia_SIGGRAPH Asia 2024_sess105_papers_854@linklings.com
SUMMARY:MoA: Mixture-of-Attention for Subject-Context Disentanglement in P
 ersonalized Image Generation
DESCRIPTION:Kuan-Chieh Wang, Daniil Ostashev, Yuwei Fang, Sergey Tulyakov,
  and Kfir Aberman (Snap Inc.)\n\nWe introduce a new architecture for perso
 nalization of text-to-image diffusion models, coined Mixture-of-Attention 
 (MoA). Inspired by the Mixture-of-Experts mechanism utilized in large lang
 uage models (LLMs), MoA distributes the generation workload between two at
 tention pathways: a personalized branch and a non-personalized prior branc
 h.\nMoA is designed to retain the original model's prior by fixing its att
 ention layers in the prior branch, while minimally intervening in the gene
 ration process with the personalized branch that learns to embed subjects 
 in the layout and context generated by the prior branch.\nA novel routing 
 mechanism manages the distribution of pixels in each layer across these br
 anches to optimize the blend of personalized and generic content creation.
  \nOnce trained, MoA facilitates the creation of high-quality, personalize
 d images featuring multiple subjects with compositions and interactions as
  diverse as those generated by the original model.\nCrucially, MoA enhance
 s the distinction between the model's pre-existing capability and the newl
 y augmented personalized intervention, thereby offering a more disentangle
 d subject-context control that was previously unattainable.\n\nRegistratio
 n Category: Full Access, Full Access Supporter\n\nLanguage Format: English
  Language\n\nSession Chair: Kfir Aberman (Decart AI)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_854&sess=sess105
END:VEVENT
END:VCALENDAR