Introduction

Earlier this year NVIDIA rolled out DLSS 4.5—the big headlines were Dynamic Multi Frame Generation, 6X Multi Frame Generation and a second-generation transformer model for Super Resolution. One piece was still missing from that release: Ray Reconstruction, which continued to use the first-generation transformer model that shipped with DLSS 4. Today NVIDIA is closing that gap with the release of DLSS 4.5 Ray Reconstruction, which brings with it an improved denoiser aimed at everything that uses ray tracing or path tracing.

Ray Reconstruction was first introduced with DLSS 3.5 in 2023 (our review), and the idea behind it hasn't changed since. During ray tracing, only limited number of rays is cast per frame, far fewer than there are pixels on your screen, which means this output alone would be full of noise. Traditionally, a hand-tuned denoiser cleans up that noise, and only once that is complete, the upscaler runs on this output—at that point all the information from the game engine about rays, light and materials is gone. Ray Reconstruction replaces the classic denoiser with an AI-powered model that is trained for this specific purpose, and it merges denoising and upscaling into a single pass. This pass has access to a lot of more data, too, so it can produce a better result (at least theoretically).



What's new today is the model behind it. According to NVIDIA, the second-generation transformer model processes 20% more parameters and delivers 35% more compute capability than the model it replaces, while landing at roughly the same performance cost. Just to confirm, both are transformer models. NVIDIA also mentions that it was trained on a larger dataset, has deeper spatial awareness of the scene, makes smarter use of the engine's pixel sampling and motion data, and gives developers finer control over temporal sampling.

Just like with DLSS 4.5 Super Resolution, you do not have to wait for game developers to patch anything. The new model is available in the NVIDIA App as driver-level override—a mechanism that works really well. As alternative method, like when you're not using the NVIDIA App, you may replace the DLL manually, we're hosting the latest NVIDIA DLSS DLL files in our Downloads Section.

In this review we'll take a look at what the new model actually changes on screen, in several games and what it costs in terms of performance.

DLSS 4 vs. DLSS 4.5 Ray Reconstruction

I think it's important to make it clear what didn't change with DLSS 4.5 Ray Reconstruction. NVIDIA still uses a transformer network, but second-generation. They did not raise the hardware requirement at all! DLSS 4.5 Ray Reconstruction runs on every GeForce RTX GPU, all the way back to the RTX 20-series—no RTX 50 exclusivity, no RTX 40 minimum, no new silicon requirement of any kind. It's all running on the existing Tensor Cores of any RTX generation.

I'm pointing this out, because the rest of DLSS 4.5 does not work that way: Multi Frame Generation and Dynamic Multi Frame Generation are RTX 50-series exclusives, and Frame Generation needs at least an RTX 40 card.