BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Hong_Kong
X-LIC-LOCATION:Asia/Hong_Kong
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:HKT
DTSTART:19911015T033000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T030656Z
LOCATION:Meeting Room S423+S424\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251218T132000
DTEND;TZID=Asia/Hong_Kong:20251218T133100
UID:siggraphasia_SIGGRAPH Asia 2025_sess155_papers_1615@linklings.com
SUMMARY:BlobCtrl: Taming Controllable Blob for Element-level Image Editing
DESCRIPTION:Yaowei Li (Peking University); Lingen Li (Chinese University o
 f Hong Kong); Zhaoyang Zhang, Xiaoyu Li, and Guangzhi Wang (Tencent); Hong
 xiang Li (The Hong Kong University of Science and Technology); Xiaodong Cu
 n (GVC Lab, Great Bay University); Ying Shan (Tencent); and Yuexian Zou (P
 eking University)\n\nAs user expectations for image editing continue to ri
 se, the demand for flexible, fine-grained manipulation of specific visual 
 elements presents a challenge for current diffusion-based methods.\n\n    
 In this work, we present BlobCtrl, a framework for element-level image edi
 ting based on a probabilistic blob-based representation. Treating blobs as
  visual primitives, BlobCtrl disentangles layout from appearance, affordin
 g fine-grained, controllable object-level elements manipulation.\n\n    Ou
 r key contributions are twofold: 1) an in-context dual-branch diffusion mo
 del that separates foreground and background processing, incorporating blo
 b representations to explicitly decouple layout and appearance; and 2) a s
 elf-supervised disentangle-then-reconstruct training paradigm with an iden
 tity-preserving loss function, along with tailored strategies to efficient
 ly leverage blob-image pairs.\n\n    To foster further research, we introd
 uce BlobData for large-scale training, and BlobBench, a benchmark for syst
 ematic evaluation. Experimental results demonstrate that BlobCtrl achieves
  state-of-the-art performance in a variety of element-level editing tasks—
 such as object addition, removal, scaling, and replacement—while maintaini
 ng computational efficiency.\n\nRegistration Category: Full Access, Full A
 ccess Supporter\n\nSession Chair: Ali Mahdavi-Amiri (Simon Fraser Universi
 ty, MARZ VFX)\n\n
END:VEVENT
END:VCALENDAR