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 B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241204T133400
DTEND;TZID=Asia/Tokyo:20241204T134600
UID:siggraphasia_SIGGRAPH Asia 2024_sess116_papers_480@linklings.com
SUMMARY:TurboEdit: Text-Based Image Editing Using Few-Step Diffusion Model
 s
DESCRIPTION:Gilad Deutch (Tel Aviv University); Rinon Gal (Tel Aviv Univer
 sity, NVIDIA Research); and Daniel Garibi, Or Patashnik, and Daniel Cohen-
 Or (Tel Aviv University)\n\nDiffusion models have opened the path to a wid
 e range of text-based image editing frameworks. However, these typically b
 uild on the multi-step nature of the diffusion backwards process, and adap
 ting them to distilled, fast-sampling methods has proven surprisingly chal
 lenging. Here, we focus on a popular line of text-based editing frameworks
  - the "edit-friendly" DDPM-noise inversion approach. We analyze its appli
 cation to fast sampling methods and categorize its failures into two class
 es: the appearance of visual artifacts, and insufficient editing strength.
  We trace the artifacts to mismatched noise statistics between inverted no
 ises and the expected noise schedule, and suggest a shifted noise schedule
  which corrects for this offset. To increase editing strength, we propose 
 a pseudo-guidance approach that efficiently increases the magnitude of edi
 ts without introducing new artifacts. All in all, our method enables text-
 based image editing with as few as three diffusion steps, while providing 
 novel insights into the mechanisms behind popular text-based editing appro
 aches.\n\nRegistration Category: Full Access, Full Access Supporter\n\nLan
 guage Format: English Language\n\nSession Chair: Dani Lischinski (Hebrew U
 niversity of Jerusalem, Google)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_480&sess=sess116
END:VEVENT
END:VCALENDAR