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:20241204T132300
DTEND;TZID=Asia/Tokyo:20241204T133400
UID:siggraphasia_SIGGRAPH Asia 2024_sess116_papers_479@linklings.com
SUMMARY:Consolidating Attention Features for Multi-view Image Editing
DESCRIPTION:Or Patashnik (Tel Aviv University); Rinon Gal (Tel Aviv Univer
 sity, NVIDIA Research); Daniel Cohen-Or (Tel Aviv University); and Jun-Yan
  Zhu and Fernando De La Torre (Carnegie Mellon University)\n\nLarge-scale 
 text-to-image models enable a wide range of image editing techniques, usin
 g text prompts or even spatial controls. However, applying these editing m
 ethods to multi-view images depicting a single scene leads to 3D-inconsist
 ent results. In this work, we focus on spatial control-based geometric man
 ipulations and introduce a method to consolidate the editing process acros
 s various views. We build on two insights: (1) maintaining consistent feat
 ures throughout the generative process helps attain consistency in multi-v
 iew editing, and (2) the queries in self-attention layers significantly in
 fluence the image structure. Hence, we propose to improve the geometric co
 nsistency of the edited images by enforcing the consistency of the queries
 . To do so, we introduce QNeRF, a neural radiance field trained on the int
 ernal query features of the edited images. Once trained, QNeRF can render 
 3D-consistent queries, which are then softly injected back into the self-a
 ttention layers during generation, greatly improving multi-view consistenc
 y. We refine the process through a progressive, iterative method that bett
 er consolidates queries across the diffusion timesteps. We compare our met
 hod to a range of existing techniques and demonstrate that it can achieve 
 better multi-view consistency and higher fidelity to the input scene. Thes
 e advantages allow us to train NeRFs with fewer visual artifacts, that are
  better aligned with the target geometry.\n\nRegistration Category: Full A
 ccess, Full Access Supporter\n\nLanguage Format: English Language\n\nSessi
 on Chair: Dani Lischinski (Hebrew University of Jerusalem, Google)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_479&sess=sess116
END:VEVENT
END:VCALENDAR