AMD researchers are exploring a new way to use generative AI in game graphics: instead of only reconstructing higher-resolution pixels, an AI model can help generate indirect lighting.
The research, titled “Temporally stable generative illumination with a one-step diffusion model,” demonstrates how a neural model can create global-illumination information from directly lit frames.
Why global illumination is expensive
Realistic light does more than travel directly from a lamp to an object. It bounces between surfaces, picking up color and illuminating areas indirectly. Simulating those interactions accurately in real time requires significant computing power.
AMD’s experiment asks whether a generative model can learn to produce that indirect-lighting result instead.
The model has to stay stable between frames
Generating one attractive frame is not enough for a game. Small inconsistencies between consecutive frames appear as distracting flicker. The researchers use a temporal VAE to improve consistency over time.
This is research, not a feature you can enable today
The current method is far too slow for mainstream real-time gaming. PC Gamer reports roughly 0.29 seconds per frame on an RTX 3090 with around 8–9GB of VRAM usage. At that speed, it is a research demonstration rather than a shipping graphics option.
Why it still matters
Upscaling and frame generation were only the beginning of AI-assisted graphics. Nvidia has been moving toward neural rendering, and AMD’s research shows that lighting itself may become another part of the graphics pipeline handled by learned models.
If hardware and algorithms become fast enough, future GPUs and consoles could use AI not merely to sharpen a rendered image but to help create parts of that image.
Source: AMD research summarized by PC Gamer, September 17, 2026.