🤖 AI Summary
To address slow convergence and high noise in real-time path tracing—leading to insufficient frame rates—this paper proposes a dynamic foveated progressive rendering framework grounded in human visual perception. Methodologically, it pioneers the tight integration of gaze prediction with path tracing, enabling adaptive undersampling in peripheral regions; it also establishes the first scalable retinal rendering baseline system built on OptiX 7.5 and CUDA 12.1. Key contributions include: (1) a global acceleration mechanism that preserves perceptually lossless image quality, and (2) a structured error-mapping evaluation framework supporting multi-resolution and multi-sample configurations. Experiments demonstrate a 25× speedup on complex scenes, with no statistically significant degradation in subjective visual quality and sustained real-time performance (≥30 FPS).
📝 Abstract
Path tracing is one of the most widespread rendering techniques for high-end graphics fidelity. However, the slow convergence time and presence of intensive noises make it infeasible for numerous real-time applications where physically corrected photorealistic effects are salient. Additionally, the increased demand for pixel density, geometric complexity, advanced material, and multiple lights hinder the algorithm from attaining an interactive frame rate for real-time applications. To address these issues, we developed a framework to accelerate path tracing through foveated rendering, a robust technique that leverages human vision. Our dynamic foveated path-tracing framework integrates fixation data and selectively lowers the rendering resolution towards the periphery. The framework is built on NVIDIA's OptiX 7.5 API with CUDA 12.1, serving as the base of future foveated path tracing research. Through comprehensive experimentation, we demonstrated the effectiveness of our framework in this paper. Depending on the scene complexity, our solution can significantly enhance rendering performance up to a factor of 25 without any notable visual differences. We further evaluated the framework using a structured error map algorithm with variable sample numbers and foveated area size.