GDC 2025 Session Viewer

On July 10th, Once Human was launched, which features a vast 16km open world with time-of-day cycles(TOD) and real-time construction of various buildings within the terrain. Consequently, the game requires a high quality pre-baked GI that supports TOD, as well as a dynamic GI that updates in real time with the player's building activities, which presents considerable challenges regarding memory usage, package size, and performance. In our talk, we will demonstrate how we address these challenges using a hybrid GI solution. For pre-baked GI, we will show how we have harnessed the power of neural networks to encode GI probe data across various times of day and player's dynamic flashlight. We also showcase a novel, machine learning-based method to automate the resolution of indoor-outdoor light leaks for probes. For real-time GI, we will show how we improve the visual effect and performance across all devices compared to previous methods.

Takeaway

Attendees can learn about a hybrid global illumination(GI) solution across various platforms(pc/mobile), how the neural network can be used to compress GI for different time frames and dynamic flashlight, and how to use machine learning to help solve the light leaking problem of GI probes, and get a deeper dive into how to address challenges for real-time GI such as memory usage, performance, and visual effect

Intended Audience

Graphic and rendering programmers, or anyone interested in global illumination. Prior knowledge about global illumination and some basic understanding of machine learning will be useful.