Temporal Gradient-Domain Path Tracing

摘要

We present a novel approach to improve temporal coherence inrnMonte Carlo renderings of animation sequences. Unlike other approachesrnthat exploit temporal coherence in a post-process, ourrntechnique does so already during sampling. Building on previousrngradient-domain rendering techniques that sample finite differencesrnover the image plane, we introduce temporal finite differencesrnand formulate a corresponding 3D spatio-temporal screened Poissonrnreconstruction problem that is solved over windowed batchesrnof several frames simultaneously. We further extend our approachrnto include second order, mixed spatio-temporal differences, an improvedrntechnique to compute temporal differences exploiting motionrnvectors, and adaptive sampling. Our algorithm can be built onrna gradient-domain path tracer without large modifications. In particular,rnwe do not require the ability to evaluate animation pathsrnover multiple frames. We demonstrate that our approach effectivelyrnreduces temporal flickering in animation sequences, significantlyrnimproving the visual quality compared to both path tracingrnand gradient-domain rendering of individual frames.