#### Articles read this week - [[Approximating dynamic global illumination in image space]] - [[Fast Global Illumination Approximations on Deep G-Buffers]] - [[Neural Global Illumination Interactive Indirect Illumination Prediction Under Dynamic Area Lights]] - [[Deep Shading Convolutional Neural Networks for Screen-Space Shading]] ---- #### New articles found - [[Deep Shading Convolutional Neural Networks for Screen-Space Shading]] - [[Screen Space Indirect Lighting with Visibility Bitmask]] --- ### Notes The early ones are around the same idea of using depth buffers as geometry information and sample around them. Calculating bounce lights from them is just introducing another layer of depth (more geometry). Neural network ones provide different inputs to models. [[Neural Global Illumination Interactive Indirect Illumination Prediction Under Dynamic Area Lights]] is trained their model for one scene and tested the results on the same scene (different camera angles and area light positions/sizes), but they are ineffective against different camera angles that model is not trained on. It uses CNN to encode scene parameters as a buffer (neural g buffer as they call), and feed this and some material information (same scene with different pre-selected materials) to a MLP model. [[Deep Shading Convolutional Neural Networks for Screen-Space Shading]] is trained for estimating different kinds of screen space techniques, it has a custom model architecture with g-buffer like information feed into it (don't have any interesting input type, and used Phong shading as reference). The resulting models still didn't quite there for executing real time, may look into more performant ones.