AMD FSR Redstone pushes the FidelityFX Super Resolution package for upscaling and frame generation into the machine learning era. Enhanced upscaling, frame generation, Ray Regeneration, and Radiance Caching all utilize ML on Radeon RX 9000 cards with the RDNA 4 architecture. We’ll look at what FSR Redstone adds compared to previous FSR 1–3 generations, how it works, what its limitations are, and how PC game support will shape up.
FSR Ray Regeneration focuses on ray-traced lighting and reflections. At its core it is a denoiser (an algorithm for removing noise in the image) using a neural network that AMD has trained specifically for ray tracing and path tracing.
When a game uses denoising algorithms, it does not need to use very high ray counts per pixel. It can afford a lower sample count, meaning an input image with significantly more noise. Denoisers are a common component of ray-traced games.
Ray Regeneration differs from these common approaches in that it uses a trained neural network capable of reconstructing a lower-sample image with higher quality and producing a cleaner and more detailed result.
The algorithm works with multiple buffers: besides the raw ray-tracing output it also considers normals and depth, diffuse and specular components, radiance, and possibly visibility information of light sources. From this data it builds a context of the scene, its materials, and its lighting, and processes it in a layer AMD refers to in its materials as the warp module—a combination of a neural network and spatial filters.
The output is a frame with filtered noise and cleaner reflections, more stable lighting, and fewer artifacts even at low ray-per-pixel counts. In practice this means the game approaches “clean” ray-tracing quality but with lower performance requirements. As a result, the number of rays per pixel can be reduced and ray-tracing effects can be optimized even for weaker GPUs.
Ray Regeneration runs directly in the engine as part of the rendering pipeline, allowing development studios to adapt it to their own pipelines while still taking advantage of AMD’s ready-made ML component.
The first implementation appeared in the newly released Call of Duty: Black Ops 7. Unlike the competing Ray Reconstruction technology, however, it is not part of the FSR Upscaling pipeline and functions independently from it.
We analyzed the impact of its settings on performance and image quality in the article CoD: Black Ops 7: RX 9060 XT, RX 9070 XT and Ray Regeneration. I will borrow a few examples from it, and you can find many more in the article itself.
RT Off (SSR High) | RT High + RT Denoiser Default | RT High + RT Ray Regeneration
A drawback of non-AI denoisers used in games is that if a reflection becomes occluded and then revealed again, details in those areas are missing and only appear as more samples are accumulated. Instead of a detailed reflection, the previously occluded region only shows a blurry patch. In motion this can be seen in the following video—in the lower half you can notice how, after the movement stops, the reflection on the water surface gradually sharpens as more samples are computed. With Ray Regeneration, only a narrow strip around the weapon shows similar issues.
The footage is slowed down from 120 to 30 frames per second and the crop is magnified 2×—mainly because of YouTube video compression. Even on monitors with a resolution of 1920 × 1080 pixels, I recommend switching the playback quality to 4K.
Ray Regeneration is currently supported in only one game, but more are expected to follow.
