©
To celebrate the launch of NVIDIA DLSS 3.5 in Cyberpunk 2077: Phantom Liberty, Digital Foundry hosted a roundtable video chat with Brian Catanzaro (VP of Applied Deep Learning Research at NVIDIA) and Jakub Knapik (VP of Art and Global Art). director of CD PROJEKT RED).
Discussing the new technology, NVIDIA’s Brian Catanzaro said that DLSS 3.5 is not only prettier than native rendering, but in some ways its frames are more realistic when combined with path tracing than native rendering with a traditional raster approach.
You know, I actually think that DLSS 3.5 makes Cyberpunk 2077 even more beautiful than native rendering. This is my belief. The reason for this, again, is that AI is able to make smarter decisions about how to render a scene than what we knew how to do without AI. I think this will continue to evolve.
Cyberpunk 2077 frames using DLSS (including frame generation) are much more “real” than traditional graphics frames. If you think about all the graphics tricks, all the occlusions, shadows, artificial reflections, screen space effects… The raster in general is just a bag of falsehood, right? So we can scrap that and start tracing the path and we can actually get real shadows and real reflections.
The only way to do this is to synthesize many pixels using AI. It would be too much work to render the path trace without some tricks. So we’re changing the methods we use, and I think we’ll end up getting more real pixels with DLSS 3.5 than without it.
CDPR’s Jakub Knapik agreed, calling rasterization “a bunch of hacks stacked on top of each other.”
It’s strange to say this, but I agree with you. This is a very interesting point of view, and you ask: what is the trade-off here? On the one hand, you have an approach to rasterization that is a bunch of hacks. You have multiple rendering layers that don’t balance each other, you just layer them on top of each other to create frames. Each individual layer is a hack, an approximation of reality: reflections in screen space and all that, and you create those pixels, as opposed to a more precise definition of reality with path tracing, where you generate something with DLSS. It used to be that by using DLSS you were sacrificing some quality for performance, but with DLSS 3.5 the image certainly looks better than without it.
The discussion of “fake frames” first arose in relation to the generation of DLSS frames, as it generates a single frame independent of the rendering pipeline and inserts a “fake frame” after each “real frame”. However, the performance boost provided by generated frames is undeniably large, especially in CPU-intensive games where a typical upscaler can do very little. It’s no coincidence that AMD is following suit with FSR 3, which will soon debut in Forspoken and Immortals of Aveum.
Specifically for DLSS 3.5, the new Ray Reconstruction feature provides greater quality benefits when scaling and ray tracing are active.
