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 S426+S427\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251216T135300
DTEND;TZID=Asia/Hong_Kong:20251216T140400
UID:siggraphasia_SIGGRAPH Asia 2025_sess122_papers_1766@linklings.com
SUMMARY:Adaptive Neural Kernels for Gradient-domain Rendering
DESCRIPTION:Matthieu Josse (Ecole Polytechnique), Joey Litalien (Independe
 nt), and Adrien Gruson (École de technologie supérieure (ÉTS))\n\nMonte Ca
 rlo methods are a cornerstone of physics-based light transport simulations
 , valued for their ability to produce high-quality photorealistic images. 
 These stochastic methods often suffer from variance, resulting in undesira
 ble noise in the rendered images. Gradient-domain rendering (GDR) techniqu
 es mitigate this problem by estimating unbiased image-space gradients via 
 so-called shift-mapping operators. While these mappings are computationall
 y efficient, they can yield high-variance gradients---and thus poor recons
 truction quality---when applied to pixels with wildly different integrals.
  We tackle this challenge by dynamically selecting the optimal set of neig
 hboring pixels for applying shift-mapping under random sequence replay. Ke
 y to our approach is a differentiable sorting network that softly ranks th
 e output of a convolutional neural network conditioned on input sample fea
 tures for weighted reconstruction. This module is carefully rigidified ove
 r time to converge to a hard top-$k$ selection, allowing end-to-end optimi
 zation with respect to the reconstruction error. Our method is versatile a
 nd can be jointly optimized with other adaptive sampling strategies. We de
 monstrate variance reduction over other traditional adaptive gradient-doma
 in methods across scenes of varying radiometric complexity.\n\nRegistratio
 n Category: Full Access, Full Access Supporter\n\nSession Chair: Elena Gar
 ces (Adobe, Universidad Rey Juan Carlos)\n\n
END:VEVENT
END:VCALENDAR