AMD published research on September 9, 2026, describing an experimental method that generates screen-space global illumination with a one-step diffusion model. A temporal decoder uses information from earlier frames to steady the lighting. The reported RTX 3090 result took 0.29 seconds per 512 × 512 GI frame, so it measures the illumination step—not a complete game-rendering pipeline. AMD presents the method as research, not a released Radeon feature.
What global illumination means
Global illumination (GI) describes how light bounces around a scene and creates indirect effects, such as color spreading from one surface to another. Ray tracing and path tracing are computational approaches that can estimate those effects; GI itself is the lighting being calculated.
AMD’s research takes a different route: it treats screen-space GI as an image-generation task. The model uses scene and lighting information to generate an indirect-lighting result, which AMD compares with a path-traced reference.
How AMD’s model generates indirect light
A one-step latent diffusion model starts with a direct-lighting frame as a hint. It also receives scene and lighting cues, including surface normals, albedo, roughness, metallic properties, noisy GI samples and radiance cues. T2I-Adapter modules feed these conditions into multiple levels of the model’s U-Net.
To keep the output steady as a scene changes, a temporal variational autoencoder (VAE) decoder receives the current latent representation and a prior result reprojected with motion vectors. It can reuse information from similar pixels in earlier frames, helping suppress flicker while retaining detail.
Training and evaluation
AMD trained the model on synthetic interior scenes rendered with Blender Cycles and evaluated it on scenes the model had not seen during training. A reported account of the training setup describes 31,000 frames, 1–30 lights per scene and 50 passes through the dataset on AMD Instinct MI210 GPUs.
AMD reports that its results on unseen synthetic scenes were visually closer to reference GI than selected comparison methods. It also reports leading results across its evaluated scenes on PSNR, SSIM, LPIPS and FovVideoVDP, measures used to assess image quality and temporal stability.
A compact variant called Ours-C uses signals that can be recovered by inverse rendering real-world images. AMD evaluated this variant on real photos; that work is distinct from deploying the method in a consumer game.
What the reported performance means
The prototype’s reported inference time on an NVIDIA GeForce RTX 3090 was 0.29 seconds per 512 × 512 GI frame, with approximately 8–9 GB of VRAM used for the GI step. That latency corresponds to about 3.45 frames per second for this step. It is not a full game’s frame rate.
AMD’s work remains research
AMD says it presented the research at ECCV 2026. Its September 9 article describes an experimental method, not a Radeon, FSR or game release.
