Paired Image to Image Translation for Ambient Occlusion Approximation

Screen space ambient occlusion seeks to accurately reproduce global illumination in real-time, thus rendering 3D scene details with smooth shadows. However, this method uses G-Buffers, which require multiple buffers to store the necessary information, which leads to some constraints, particularly regarding performance and memory management. This paper proposes a novel approach inspired by Screen space ambient occlusion, but using a single map. Moreover, ambient occlusion can be realistically simulated using generative models, such as image-to-image translation using deep neural networks. In this framework, we created a dataset of 10,000 synthetic images and developed a model to generate ambient occlusion efficiently. The suggested approach generates images similar to those created by classical methods, without the need for preprocessing or computation in the 3D space of the scene. The experimental results highlight the ability of our technique to generate high-quality and high-precision images.