Research on global illumination, a technique for simulating light behavior in computer graphics that reproduces realistic lighting effects by considering the influence of indirect light within a scene, is actively progressing. Within this field, this paper proposes a focused sampling method for efficiently rendering circle of confusion and depth of field effects that account for the characteristics of camera lenses. Traditional Monte Carlo methods attempting to reproduce these effects have employed perturbation techniques, such as randomly distributing rays on the lens surface to spread rays passing through the lens in different directions. However, uniform lens sampling suffers from the problem of inadequate sampling in critical areas like highlight regions within the scene, resulting in excessive noise. While techniques called path guiding can learn more important incident light intensity distributions and improve sampling efficiency, no path guiding method specifically for reproducing circle of confusion and depth-of-field effects has been proposed to date. This is primarily due to the parallax problem: even for adjacent pixels, a slight difference in position on the lens can drastically change the position they reach within the scene.
Therefore, we propose a method that introduces importance sampling in lens space and concentrates rays toward the highlight spot. This method models the highlight spot in space using a light intensity field and guides camera ray generation by transforming it into lens space via bipolar-cone projection. In the implementation, we represent the highlight spot in world coordinates using a 3D Gaussian distribution function, then transform it into lens coordinates to absorb the effects of parallax. This allows us to effectively utilize the importance distribution of lens sampling even under strong parallax. We achieved a fast and robust guidance method. Furthermore, we provide a method suitable for production rendering, capable of generating large volumes of high-quality images for final delivery in films, commercials, games, etc., while minimizing computational load and memory overhead. Experimental results demonstrate that our proposed method significantly improves rendering efficiency for circle of confusion and depth-of-field effects compared to conventional uniform sampling or resampling methods.