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 B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241204T130000
DTEND;TZID=Asia/Tokyo:20241204T131100
UID:siggraphasia_SIGGRAPH Asia 2024_sess116_papers_308@linklings.com
SUMMARY:DiffUHaul: A Training-Free Method for Object Dragging in Images
DESCRIPTION:Omri Avrahami (Hebrew University of Jerusalem), Rinon Gal (Tel
  Aviv University), Gal Chechik (NVIDIA), Ohad Fried (The Interdisciplinary
  Center Herzliya), Dani Lischinski (Hebrew University of Jerusalem), and A
 rash Vahdat and Weili Nie (NVIDIA)\n\nText-to-image diffusion models have 
 proven effective for solving many image editing tasks.\n    However, the s
 eemingly straightforward task of seamlessly relocating objects within a sc
 ene remains surprisingly challenging. Existing methods addressing this pro
 blem often struggle to function reliably in real-world scenarios due to la
 cking spatial reasoning. \n    In this work, we propose a training-free me
 thod, dubbed \emph{DiffUHaul}, that harnesses the spatial understanding of
  a \emph{localized} text-to-image model, for the object dragging task.\n  
   Blindly manipulating layout inputs of the localized model tends to cause
  low editing performance due to the intrinsic entanglement of object repre
 sentation in the model. To this end, we first apply attention masking in e
 ach denoising step to make the generation more disentangled across differe
 nt objects and adopt the self-attention sharing mechanism to preserve the 
 high-level object appearance. Furthermore, we propose a new diffusion anch
 oring technique: in the early denoising steps, we interpolate the attentio
 n features between source and target images to smoothly fuse new layouts w
 ith the original appearance; in the later denoising steps, we pass the loc
 alized features from the source images to the interpolated images to retai
 n fine-grained object details. To adapt DiffUHaul to real-image editing, w
 e apply a DDPM self-attention bucketing that can better reconstruct real i
 mages with the localized model.\n    Finally, we introduce an automated ev
 aluation pipeline for this task and  showcase the efficacy of our method. 
 Our results are reinforced through a user preference study.\n\nRegistratio
 n Category: Full Access, Full Access Supporter\n\nLanguage Format: English
  Language\n\nSession Chair: Dani Lischinski (Hebrew University of Jerusale
 m, Google)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_308&sess=sess116
END:VEVENT
END:VCALENDAR