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 S221\, Level 2
DTSTART;TZID=Asia/Hong_Kong:20251216T140400
DTEND;TZID=Asia/Hong_Kong:20251216T141500
UID:siggraphasia_SIGGRAPH Asia 2025_sess123_papers_1905@linklings.com
SUMMARY:MILO: A Lightweight Perceptual Quality Metric for Image and Latent
 -Space Optimization
DESCRIPTION:Ugur Cogalan, Mojtaba Bemana, Karol Myszkowski, Hans-Peter Sei
 del, and Colin Groth (Max Planck Institute for Informatics)\n\nWe present 
 MILO (Metric for Image- and Latent-space Optimization), a lightweight, mul
 tiscale, perceptual metric for full-reference image quality assessment (FR
 -IQA).\nMILO is trained using pseudo-MOS (Mean Opinion Score) supervision,
  in which reproducible distortions are applied to diverse images and score
 d via an ensemble of recent quality metrics that account for visual maskin
 g effects.\nThis approach enables accurate learning without requiring larg
 e-scale human-labeled datasets. Despite its compact architecture, MILO out
 performs existing metrics across standard FR-IQA benchmarks and offers fas
 t inference suitable for real-time applications.\nBeyond quality predictio
 n, we demonstrate the utility of MILO as a perceptual loss in both image a
 nd latent domains. In particular, we show that spatial masking modeled by 
 MILO, when applied to latent representations from a VAE encoder within Sta
 ble Diffusion, enables efficient and perceptually aligned optimization. By
  combining spatial masking with a curriculum learning strategy, we first p
 rocess perceptually less relevant regions before progressively shifting th
 e optimization to more visually distorted areas. This strategy leads to si
 gnificantly improved performance in tasks like denoising, super-resolution
 , and face restoration, while also reducing computational overhead.\nMILO 
 thus functions as both a state-of-the-art image quality metric and as a pr
 actical tool for perceptual optimization in generative pipelines.\n\nRegis
 tration Category: Full Access, Full Access Supporter\n\nSession Chair: Suy
 eon Choi (Stanford University)\n\n
END:VEVENT
END:VCALENDAR
