BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260805T205451Z
LOCATION:Room 408 B
DTSTART;TZID=America/Los_Angeles:20260722T090000
DTEND;TZID=America/Los_Angeles:20260722T103000
UID:siggraph_SIGGRAPH 2026_sess154@linklings.com
SUMMARY:Image Generation
DESCRIPTION:On-the-fly Repulsion in the Contextual Space for Rich Diversit
 y in Diffusion Transformers\n\nModern T2I DiTs are powerful yet often lack
  diversity. We identify a novel intervention point within DiTs: the Contex
 tual Space at the transformer block level, where textual tokens are enrich
 ed by image content. We propose a method that repels a batch of samples in
  this space to produce diverse gene...\n\n\nOmer Dahary (Tel Aviv Universi
 ty, Snap Research); Benaya Koren (Tel Aviv University); and Daniel Garibi 
 and Daniel Cohen-Or (Tel Aviv University, Snap Research)\n----------------
 -----\nBFS: Back-to-Front Layered Image Synthesis via Knowledge Transfer\n
 \nBFS is a novel generation-based framework for layered image synthesis th
 at produces foreground layers with realistic visual effects while harmoniz
 ing with a given background. Using a dual-branch diffusion model and two-s
 tage training with unlayered data, it overcomes data scarcity and achieves
  super...\n\n\nKyoungkook Kang (Samsung) and Gyujin Sim and Sunghyun Cho (
 POSTECH)\n---------------------\nColorful-Noise: Training-Free Low-Frequen
 cy Noise Manipulation for Color-Based Conditional Image Generation\n\nColo
 rful-Noise is a training-free image generation conditioning method that le
 verages FFT noise frequency manipulation. We show that low frequencies dom
 inate output appearance and can encode structure from scribbles, color map
 s, or reference images — enabling anyone to condition image generati...\n\
 n\nNadav Z. Cohen, Ofir Abramovich, and Ariel Shamir (Reichman University)
 \n---------------------\nInspiration Seeds: Learning Non-Literal Visual Co
 mbinations for Generative Exploration\n\nInspiration Seeds is a generative
  framework that combines two input images into diverse, visually coherent 
 compositions—without text prompts. Trained on synthetic triplets using CLI
 P Sparse Autoencoders, it uncovers latent visual relationships between inp
 uts, supporting early-stage creative ex...\n\n\nKfir Goldberg (Bria AI), E
 lad Richardson (Runway), and Yael Vinker (Massachusetts Institute of Techn
 ology (MIT))\n---------------------\nCanvas-to-Image: Compositional Image 
 Generation with Multimodal Controls\n\nCanvas-to-Image is a unified framew
 ork enhancing diffusion models' compositional and multimodal control throu
 gh a single canvas interface. By encoding subject references, bounding box
 es, and poses into one composite canvas, it enables integrated visual-spat
 ial reasoning. This approach streamlines he...\n\n\nYusuf Dalva (Virginia 
 Tech University); Gordon Guocheng Qian and Maya Goldenberg (Snap); Tsai-Sh
 ien Chen (Snap, University of California Merced); Kfir Aberman and Sergey 
 Tulyakov (Snap); Pinar Yanardag (Virginia Tech University); and Kuan-Chieh
  Jackson Wang (Snap)\n---------------------\nComboStoc: Combinatorial Stoc
 hasticity for Diffusion Generative Models\n\nWe study an important factor 
 in diffusion generative models: combinatorial complexity. Existing trainin
 g schemes insufficiently sample the joint space of dimensions and attribut
 es, leading to weaker test time performance. We address this with ComboSto
 c, a simple stochastic process that better exploi...\n\n\nRui Xu and Jiepe
 ng Wang (University of Hong Kong), Hao Pan (Tsinghua University), Yang Liu
  and Xin Tong (Microsoft Research Asia), Shiqing Xin and Changhe Tu (Shand
 ong University), Taku Komura (The University of Hong Kong), and Wenping Wa
 ng (Texas A&M University)\n---------------------\nImage Generation - Inter
 active Discussion\n\nAfter the summary presentations, attendees will parti
 cipate in an interactive discussion. Outside the room will be a series of 
 poster boards for authors to gather around with the audience. Authors are 
 invited to bring any material related to their paper that could instigate 
 further conversation such...\n\n\nInterest Area: Research & Education\n\nK
 eyword: Artificial Intelligence/Machine Learning, Generative AI, Modeling\
 n\nRegistration Category: Full Conference Supporter, Full Conference\n\nSe
 ssion Chair: Justin Solomon (Massachusetts Institute of Technology (MIT))
END:VEVENT
END:VCALENDAR