GoKawiilNvidia first debuted DLSS 5 earlier this year to a strong, critical response. The upscaling tech was almost slammed by some press and enthusiasts for overstepping its boundaries and going too far with AI.
At the time, Nvidia reassured the community that DLSS 5 would preserve artistic intent and CEO Jensen Huang even went as far as to claim that gamers didn't understand it. Now, at SIGGRAPH, the company has once again showcased DLSS 5, and it admittedly looks different this time around.
NVIDIA Keynote Live at SIGGRAPH 2026 - YouTube Watch On
The main takeaway is the introduction of three different models with varying levels of detail and impact on performance. A game doesn't need to be confined to a single model the whole time; instead, developers can choose to employ different models for different scenes. Individual elements of scene, such as the characters versus the world, can also be fine-tuned in respect to how much DLSS touches them up (or not at all).
Latest Videos From Watch full video here:
Then, you have the liberty to switch between the models in real time without incurring latency. DLSS 5 uses a combination of techniques to take the original rendered frame and apply effects on top, going beyond just upscaling the image. These days, DLSS, FSR, and XeSS are mandatory parts of the equation, acting as the preferred anti-aliasing solution in modern titles, and helping in other areas like ray tracing.
The upscaling part works conventionally and is still responsible for dictating the foundational blocks of the image, like lighting and geometry. The part that was so heavily criticized is therefore optional, as it only comes into play after the original rendered frame has been upscaled. However, it's unclear whether developers (or you) can actively choose to not process the frame through the beautification features.
Nvidia outlined three main challenges that the company faced when developing DLSS 5, some of which is probably reactionary after the initial reveal. The first challenge is preserving the original creative vision, which we already explained. The second challenge was to handle one frame at a time, which generative AI models don't — they process multiple frames together. That means DLSS 5 shouldn't negatively affect response times.
... continue reading