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:20260817T171530Z
LOCATION:Hall B7 (1)\, B Block\, Level 7
DTSTART;TZID=Asia/Tokyo:20241205T131400
DTEND;TZID=Asia/Tokyo:20241205T132800
UID:siggraphasia_SIGGRAPH Asia 2024_sess132_papers_249@linklings.com
SUMMARY:Monkey See, Monkey Do: Harnessing Self-attention in Motion Diffusi
 on for Zero-shot Motion Transfer
DESCRIPTION:Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, and Rotem Shale
 v-Arkushin (Tel Aviv University); Ohad Fried (Reichman University); and Am
 it Haim Bermano and Daniel Cohen-Or (Tel Aviv University)\n\nGiven the rem
 arkable results of motion synthesis with diffusion models, a natural quest
 ion arises: how can we effectively leverage these models for motion editin
 g? Existing diffusion-based motion editing methods overlook the profound p
 otential of the prior embedded within the weights of pre-trained models, w
 hich enables manipulating the latent feature space; hence, they primarily 
 center on handling the motion space. In this work, we explore the attentio
 n mechanism of pre-trained motion diffusion models. We uncover the roles a
 nd interactions of attention elements in capturing and representing intric
 ate human motion patterns, and carefully integrate these elements to trans
 fer a leader motion to a follower one while maintaining the nuanced charac
 teristics of the follower, resulting in zero-shot motion transfer. Manipul
 ating features associated with selected motions allows us to confront a ch
 allenge observed in prior motion diffusion approaches, which use general d
 irectives (e.g., text, music) for editing, ultimately failing to convey su
 btle nuances effectively. Our work is inspired by how a monkey closely imi
 tates what it sees while maintaining its unique motion patterns; hence we 
 call it Monkey See, Monkey Do, and dub it MoMo. Employing our technique en
 ables accomplishing tasks such as synthesizing out-of-distribution motions
 , style transfer, and spatial editing. Furthermore, diffusion inversion is
  seldom employed for motions; as a result, editing efforts focus on genera
 ted motions, limiting the editability of real ones. MoMo harnesses motion 
 inversion, extending its application to both real and generated motions. E
 xperimental results show the advantage of our approach over the current ar
 t. In particular, unlike methods tailored for specific applications throug
 h training, our approach is applied at inference time, requiring no traini
 ng. Our webpage, https://monkeyseedocg.github.io, includes links to videos
  and code.\n\nRegistration Category: Full Access, Full Access Supporter\n\
 nLanguage Format: English Language\n\nSession Chair: Yi Zhou (Roblox)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_249&sess=sess132
END:VEVENT
END:VCALENDAR