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 S421\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251217T172400
DTEND;TZID=Asia/Hong_Kong:20251217T173500
UID:siggraphasia_SIGGRAPH Asia 2025_sess143_papers_1660@linklings.com
SUMMARY:In-2-4D: Inbetweening from Two Single-View Images to 4D Generation
DESCRIPTION:Sauradip Nag (Simon Fraser University), Daniel Cohen-Or (Tel A
 viv University), and Hao Zhang and Ali-Mahdavi Amiri (Simon Fraser Univers
 ity)\n\nWe pose a new problem, In-2-4D, for generative 4D (i.e., 3D + moti
 on)\ninbetweening to interpolate two single-view images. In contrast to vi
 deo/4D\ngeneration from only text or a single image, our interpolative tas
 k can lever-\nage more precise motion control to better constrain the gene
 ration. Given\ntwo monocular RGB images representing the start and end sta
 tes of an\nobject in motion, our goal is to generate and reconstruct the m
 otion in 4D,\nwithout making assumptions on the object category, motion ty
 pe, length,\nor complexity. To handle such arbitrary and diverse motions, 
 we utilize\na foundational video interpolation model for motion prediction
 .\nHowever, large frame-to-frame motion gaps can lead to ambiguous\ninterp
 retations. To this end, we employ a hierarchical approach to\nidentify key
 frames that are visually close to the input states while\nexhibiting signi
 ficant motions, then generate smooth fragments be-\ntween them. For each f
 ragment, we construct a 3D representation of\nthe keyframe using Gaussian 
 Splatting (3DGS). The temporal frames\nwithin the fragment guide the motio
 n, enabling their transforma-\ntion into dynamic 3DGS through a deformatio
 n field. To improve\ntemporal consistency and refine the 3D motion, we exp
 and the self-\nattention of multi-view diffusion across timesteps and appl
 y rigid\ntransformation regularization. Finally, we merge the independentl
 y\ngenerated 3D motion segments by interpolating boundary defor-\nmation f
 ields and optimizing them to align with the guiding video,\nensuring smoot
 h and flicker-free transitions. Through extensive\nqualitative and quantit
 ive experiments as well as a user study, we\ndemonstrate the effectiveness
  of our method and design choices.\n\nRegistration Category: Full Access, 
 Full Access Supporter\n\nSession Chair: Young Min Kim (Seoul National Univ
 ersity)\n\n
END:VEVENT
END:VCALENDAR
