AMD researchers presented a lightweight attention-based neural model for real-time global illumination that reconstructs off-screen indirect lighting by combining screen-space G-buffers with reflective shadow map data from light sources. Three encoders feed a multi-head attention module whose outputs are decoded into diffuse and specular indirect illumination, composed with direct lighting. Trained on 11 synthetic scenes, the 2.19M-parameter model generalizes to unseen scenes, outperforming a prior CNN-based method (BCNN) on complex specular highlights, and runs at 45.56ms per 512x512 frame (unoptimized FP32 PyTorch) on an AMD Instinct MI250 accelerator.
Questions this post answers
How fast does AMD's attention-based indirect illumination model run on an MI250 GPU?
The unoptimized model runs in 45.56 milliseconds at 512x512 resolution on an AMD Instinct MI250 accelerator, using an FP32 PyTorch implementation without inference optimizations. The network itself is lightweight, with 2.19 million parameters and 432 GFLOPs, tested on a server with an AMD EPYC 7763 processor and 2TB RAM running Ubuntu 22.04.5 LTS. Rendering engineers benchmarking neural GI models against real-time budgets can follow this research on daily.dev.
How does reflective shadow map data help neural global illumination models see off-screen geometry?
Reflective shadow map (RSM) texels rendered from each light source's perspective are treated as indirect virtual point lights, capturing geometry outside the main camera's view frustum. Combining this RSM data with screen-space G-buffers and a noisy one-bounce buffer lets an attention-based model reconstruct off-screen indirect shadows and specular highlights that purely screen-space approaches miss. Anyone comparing off-screen GI techniques can track neural rendering research like this on daily.dev.