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:20260817T171532Z
LOCATION:Hall B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241206T153100
DTEND;TZID=Asia/Tokyo:20241206T154300
UID:siggraphasia_SIGGRAPH Asia 2024_sess149_papers_936@linklings.com
SUMMARY:Procedural Material Generation with Reinforcement Learning
DESCRIPTION:Beichen Li (MIT CSAIL, Adobe Research); Yiwei Hu, Paul Guerrer
 o, and Milos Hasan (Adobe Research); Liang Shi (MIT CSAIL); Valentin Desch
 aintre (Adobe Research); and Wojciech Matusik (MIT CSAIL)\n\nModern 3D con
 tent creation heavily relies on procedural assets. In particular, procedur
 al materials are ubiquitous in the industry, but their manipulation remain
 s challenging. Previous work conditionally generates procedural graphs tha
 t match a given input image. However, the parameter generation step limits
  how accurately the generated graph matches the input image, due to a reli
 ance on supervision with scarcely available procedural data. We propose to
  improve parameter prediction accuracy for image-conditioned procedural ma
 terial generation by leveraging reinforcement learning (RL) and present th
 e first RL approach for procedural materials. RL circumvents the limited a
 vailability of procedural data, the domain gap between real and synthetic 
 materials, and the need for end-to-end differentiable loss functions. Give
 n a target image, we retrieve a procedural material and use an RL-trained 
 transformer model to predict a set of parameters that reconstruct the targ
 et image as closely as possible. We show that using RL significantly impro
 ves parameter prediction to match a given target image compared to supervi
 sed methods on both synthetic and real target images.\n\nRegistration Cate
 gory: Full Access, Full Access Supporter\n\nLanguage Format: English Langu
 age\n\nSession Chair: Valentin Deschaintre (Adobe Research)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_936&sess=sess149
END:VEVENT
END:VCALENDAR