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:20251218T030653Z
LOCATION:Meeting Room S423+S424\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251216T133100
DTEND;TZID=Asia/Hong_Kong:20251216T134200
UID:siggraphasia_SIGGRAPH Asia 2025_sess121_papers_2027@linklings.com
SUMMARY:Harnessing Diffusion-Yielded Score Priors for Image Restoration
DESCRIPTION:Xinqi Lin (Shenzhen Institute of Advanced Technology，Chinese A
 cademy of Sciences, University of Chinese Academy of Sciences); Fanghua Yu
  (Shenzhen Institute of Advanced Technology，Chinese Academy of Sciences); 
 Jinfan Hu (Shenzhen Institutes of Advanced Technology，Chinese Academy of S
 ciences, University of Chinese Academy of Sciences); Zhiyuan You (Shenzhen
  Institutes of Advanced Technology，Chinese Academy of Sciences, The Chines
 e University of Hong Kong); Wu Shi (Shenzhen Institutes of Advanced Techno
 logy，Chinese Academy of Sciences); Jimmy S. Ren (Hong Kong Metropolitan Un
 iversity, SenseTime Research); Jinjin Gu (INSAIT, Sofia University); and C
 hao Dong (Shenzhen Institutes of Advanced Technology，Chinese Academy of Sc
 iences)\n\nDeep image restoration models aim to learn a mapping from degra
 ded image space to natural image space. However, they face several critica
 l challenges: removing degradation, generating realistic details, and ensu
 ring pixel-level consistency. Over time, three major classes of methods ha
 ve emerged, including MSE-based, GAN-based, and diffusion-based methods. H
 owever, they fail to achieve a good balance between restoration quality, f
 idelity, and speed. We propose a novel method, HYPIR, to address these cha
 llenges. Our solution pipeline is straightforward: it involves initializin
 g the image restoration model with a pre-trained diffusion model and then 
 fine-tuning it with adversarial training. This approach does not rely on d
 iffusion loss, iterative sampling, or additional adapters. We theoreticall
 y demonstrate that initializing adversarial training from a pre-trained di
 ffusion model positions the initial restoration model very close to the na
 tural image distribution. Consequently, this initialization improves numer
 ical stability, avoids mode collapse, and substantially accelerates the co
 nvergence of adversarial training. Moreover, HYPIR inherits the capabiliti
 es of diffusion models with rich user control, enabling text-guided restor
 ation and adjustable texture richness. Requiring only a single forward pas
 s, it achieves faster convergence and inference speed than diffusion-based
  methods. Extensive experiments show that HYPIR outperforms previous state
 -of-the-art methods, achieving efficient and high-quality image restoratio
 n.\n\nRegistration Category: Full Access, Full Access Supporter\n\nSession
  Chair: Paul Debevec (Eyeline, USC Institute for Creative Technologies (IC
 T))\n\n
END:VEVENT
END:VCALENDAR