AMD just published research on using AI image generation to fake global illumination, the hardest part of making a rendered scene look real.
AMD's technical staff released a paper called 'Temporally stable generative illumination with a one-step diffusion model,' describing a technique that hands indirect lighting off to a modified version of Stable Diffusion 2.1 Turbo. Rather than tracing how light bounces around a scene the way path tracing does, the model takes a frame with only direct lighting as a hint and generates the bounced light in a single denoising step. It was trained on 31,000 synthetic interior frames rendered in Blender's Cycles engine, run through the dataset 50 times on AMD Instinct MI210 GPUs. A 'temporal VAE' component uses motion vectors to keep the lighting from flickering between frames, and researcher SungYe Kim says the method recovers soft shadows and bounce lighting more faithfully than two rival techniques he named: RGB to X, which he says hallucinates content that isn't in the scene, and DiffusionRenderer, which he says distorts geometry.
This is the clearest public evidence yet that AMD's FSR Diamond neural-rendering push, announced earlier this year, is more than a keynote slide. It also reads like AMD trying to avoid a repeat of Nvidia's DLSS 5 rollout, where a splashy consumer-facing announcement got walked back into more of a developer tool after backlash.
On an RTX 3090, the model ran at roughly 3.45 frames per second at 512x512 resolution. That's promising for a research paper and nowhere near shippable in a real game, a reminder that 'coming soon' in GPU research often means years, not months.