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:20250110T023309Z
LOCATION:Hall B7 (1)\, B Block\, Level 7
DTSTART;TZID=Asia/Tokyo:20241203T131100
DTEND;TZID=Asia/Tokyo:20241203T132300
UID:siggraphasia_SIGGRAPH Asia 2024_sess105_papers_824@linklings.com
SUMMARY:ReVersion: Diffusion-Based Relation Inversion from Images
DESCRIPTION:Technical Papers\n\nZiqi Huang, Tianxing Wu, Yuming Jiang, Kel
 vin C.K. Chan, and Ziwei Liu (S-Lab for Advanced Intelligence, Nanyang Tec
 hnological University Singapore)\n\nDiffusion models gain increasing popul
 arity for their generative capabilities. Recently, there have been surging
  needs to generate customized images by inverting diffusion models from ex
 emplar images, and existing inversion methods mainly focus on capturing ob
 ject appearances (i.e., the "look"). However, how to invert object relatio
 ns, another important pillar in the visual world, remains unexplored.\nIn 
 this work, we propose the Relation Inversion task, which aims to learn a s
 pecific relation (represented as "relation prompt") from exemplar images. 
 Specifically, we learn a relation prompt with a frozen pre-trained text-to
 -image diffusion model. The learned relation prompt can then be applied to
  generate relation-specific images with new objects, backgrounds, and styl
 es. \n\nTo tackle the Relation Inversion task, we propose the ReVersion Fr
 amework.\nSpecifically, we propose a novel "relation-steering contrastive 
 learning" scheme to steer the relation prompt towards relation-dense regio
 ns, and disentangle it away from object appearances. \nWe further devise "
 relation-focal importance sampling" to emphasize high-level interactions o
 ver low-level appearances (e.g., texture, color).\nTo comprehensively eval
 uate this new task, we contribute the ReVersion Benchmark, which provides 
 various exemplar images with diverse relations. Extensive experiments vali
 date the superiority of our approach over existing methods across a wide r
 ange of visual relations. Our proposed task and method could be good inspi
 rations for future research in various domains like generative inversion, 
 few-shot learning, and visual relation detection.\n\nRegistration Category
 : Full Access, Full Access Supporter\n\nLanguage Format: English Language\
 n\nSession Chair: Kfir Aberman (Snap)
URL:https://asia.siggraph.org/2024/program/?id=papers_824&sess=sess105
END:VEVENT
END:VCALENDAR