Reproduce “Forget Superresolution, Sample Adaptively (when Path Tracing)” on OpenResearch

About this paper

Martin Bálint, Corentin Salaün, Hans-Peter Seidel, Karol Myszkowski

Researchers at the Max Planck Institute for Informatics developed an end-to-end adaptive sampling and denoising pipeline for real-time path tracing that robustly operates at sub-1-spp budgets. This method, incorporating a novel differentiable stochastic sampler and perceptual loss, consistently surpassed state-of-the-art superresolution and adaptive sampling baselines, achieving a 25.41 dB PSNR at 0.25 samples per pixel compared to DLSS 4's 22.88 dB.

Key insights

• Optimizing sample distribution adaptively, informed by perceptual metrics, yields superior visual quality in path tracing compared to uniformly reducing resolution via superresolution.
• Stable and unbiased differentiable sampling can be achieved at extreme sparsity (down to 0.01 samples per pixel) through a relaxed stochastic rounding formulation.
• Tailoring denoiser architectures, such as using gather-based filters and learned demodulation, is crucial for robustly reconstructing images from highly sparse and non-uniform inputs.

Summary by alphaXiv.