Interactive Monte-Carlo Ray-Tracing Upsampling

Abstract

We propose a practical method to approximate global illumination at interactive framerates for dynamic scenes. We address multi-bounce, visibility-aware indirect lighting, for diffuse to moderately glossy materials, relying on GPU-accelerated raytracing for this purpose. While Monte-Carlo ray-tracing algorithms offer unbiased results, they produce images which are, under interactive constraints, extremely noisy, even with GPU acceleration. Unfortunately, filtering them to reach visual appeal induces a large kernel, which is not compatible with interactive framerate. We address this problem using a simple downsampling approach. First, we trace indirect paths on a uniformly distributed subset of pixels, decorrelating diffuse and specular components of lighting. Then, we perform a joint bilateral upsampling on both components, taking inspiration from deferred shading by driving this upsampling with a full-resolution G-Buffer. Our solution provides smooth results, does not require any pre-computations, and is both easy to implement and flexible, as it can be used with any generation strategy for indirect rays.