Edge- and Eccentricity-Guided Foveated Path Tracing with Adaptive Russian Roulette

This paper presents a perceptually guided foveated path tracing framework for real-time rendering that exploits the non-uniform sensitivity of human vision. The method combines three components: (1) eccentricity-dependent sample allocation derived from a cortical magnification factor model, (2) edge-aware importance modulation that increases sampling density near high-contrast image structures, and (3) foveated Russian roulette termination that shortens expected path lengths in peripheral regions while preserving unbiasedness through throughput compensation. Evaluated on six complex scenes at 4K resolution, the approach achieves substantial performance improvements over uniform 4-SPP path tracing, reaching 5.3–6.5× higher frame rates without denoising and 2.5–3.7× higher frame rates with Intel Open Image Denoise. Perceptual evaluation using FovVideoVDP shows that the method maintains or slightly improves visual quality despite tracing approximately 2.5× fewer rays. The results demonstrate an effective quality–performance trade-off for real-time photorealistic rendering and suggest promising applications in XR systems and future gaze-contingent rendering environments.