AMD researchers have introduced a new method for real-time global illumination called Temporally Stable Generative Illumination using a one-step diffusion model. This approach aims to replace resource-intensive ray tracing by treating global illumination as an image generation task rather than a traditional computational process.
The technology is part of the upcoming FSR Diamond suite, which also features next-generation neural rendering, advanced upscaling, and multi-frame generation. The model utilizes Stable Diffusion 2.1 Turbo, taking a direct lighting frame as a prompt to generate realistic reflections and soft shadows.
The model was trained on a dataset of 31 000 pre-rendered interior scene frames, each containing between 1 and 30 light sources. Training involved 50 passes over the dataset using AMD Instinct MI210 GPUs. To ensure visual consistency and prevent flickering, the team implemented a temporal VAE that uses motion vectors to align frames.
Current performance tests on an RTX 3090 GPU show the model processing frames at a rate of 0,29 seconds per frame, or approximately 3,45 frames per second at a 512 x 512 resolution. The process consumes 8–9 GB of VRAM, indicating that the technology is currently in the experimental stage and not yet ready for consumer gaming applications.
