In Defense of DLSS 5

Furor erupted as Nvidia announced the next generation of DLSS: DLSS 5. DLSS (short for Deep Learning Super Sampling) is an image upscaling technology, primarily for video games, developed by Nvidia for their RTX graphics cards series. First released in 2018 to a limited sample of games, it works by rendering the game at a lower resolution and then using deep learning techniques to enhance the image quality. Imagine those old-school AI image upscalers, back when they were cool in the early 2020s, where you could turn a low-resolution photo into something much clearer and readable, but real-time for video games. DLSS technology would improve performance while still retaining visual appeal to the user; of course, simply rendering the game at higher graphical settings would be even better for visual appeal, but by DLSS, the game could achieve a pretty similar appearance to those higher graphical settings while being significantly less demanding on your computer.

Anyone who has used DLSS in its current form, DLSS 4, would likely tell you that the technology is fantastic. Visuals look great and high-resolution, albeit with the occasional AI artifacts that tend to occur with fast-moving visuals or idiosyncratic objects. In these cases, the model likely hadn't had much training on things to recognize these edge cases well. And one gets all of this while having significant performance benefits, and it only got better and better with passing updates and the release of the transformer model for DLSS 4.

So when Jensen Huang, on March 16, 2026 announced the release of DLSS 5, which would take even more liberties with the AI technology that DLSS is, feedback was expectedly... unexpected. Online, it has been thoroughly censured, ridiculed, and defenders of DLSS 5 (or, really, just those who haven't taken a hypercritical stance against it) have also faced hate, vitriol, death threats, and more. Philip Dyer, more commonly known as 3kliksphilip or 2kliksphilip or kliksphilip online, made a video covering DLSS 5 on his 2kliksphilip channel. He didn't praise it. He didn't criticize it. As a technological optimist, he said "it works ... [and] it's already showing some very powerful results."

So, what exactly did DLSS 5 change that has resulted in this uproar? Previous DLSS versions worked mainly by scanning areas of the screen, context-aware, and then "filling in" the gaps of the pixels, to describe it modestly. However, DLSS 5 is more baked into the game, being aware of game textures and lighting now. Now, it performs deep learning on the calculations going on within the rendering engine of the game.

Terrible!

To be less facetious, to the human eye, it seemingly applies an AI-like filter over the image in addition to the previously-existing DLSS technology, adding a bit of the "uncanny" look that many of us associate with AI-generated images. While this is a valid point of concern for video games, especially those with more unique art direction, the backlash exposes a fundamental misunderstanding of what DLSS has always been. Deep Learning Super Sampling. Notice the words "Deep Learning." From its inception in 2018, DLSS was never a simple sharpening filter like AMD's FSR, and more of Intel's XeSS: it was always a generative technology. It didn't just "find" missing pixels; at the start, with the first-ever DLSS it similarly hallucinated them into existence based on patterns it learned from high-resolution imagery. DLSS 2 and beyond looked at previous frames and used predictive motion vectors to estimate the pixels to be generated to patch up the seams. This is what has made the later versions of DLSS look so good. Now, DLSS has graduated from guessing the placement of stray pixels to guessing the intent of textures and shading and so much more. It has truly become much more intelligent. It is finally, fully, playing the game.

Yes, it doesn't look the best right now. It's a demo. Yes, they needed two 5090s to run it. And it's a demo. But I’m an optimist, like Philip. Thanks to Philip saying that the tech shown at the demo was functional, he was hanged, drawn, and quartered online. He also said that it was controversial, but I guess not many people cared enough to listen to that point. This continuous technophobia of generative AI has bled into anything relating to AI, whether it be machine learning, deep learning, neural networks, and so on. This fear of artistic intent being lost due to DLSS 5 is silly. It's like those videos comparing "soulless remake" versus the beautiful, 15-year-older-and-clearly-graphically-superior original game.

Another feature of DLSS 5: it's software-side. The game developer helps with implementing it. That fancy lighting? That upscaled wall texture? Yes, it's partially generated by AI... but it's also tuned by the game developer. They oversee and modify the technology with their eyes and intents and their minds on the art style. Furthermore, it’s still building off the existing rendered graphics in the game, not wholly re-rendering the visuals from scratch. You could, then, say that it's still not "as the artist intended" with the randomness and creativeness of the AI. The artist also likely didn't intend for you to run the game on medium graphics, or run it on a 720p resolution, or even run it with any sort of upscaling, be it DLSS 4, 3 or Intel or AMD's own upscaling technologies. The soundtrack composers likely didn't intend for you to listen to the audio on your terrible bass-boosted gaming headsets with no account for the head-related transfer function either.

The true concern is with Nvidia’s shift away from optimizing performance, opting for increasing visual fidelity as seen in the demo. I am not aware of whether or not this has been a change in Nvidia’s goal for DLSS, being to invert its original promise: performance for AI artifacts. However, that Nvidia hasn’t made a statement about it says enough. DLSS is here to improve performance. DLSS 1 upon release worsened performance at times. It certainly wasn’t easy to run in its prototype phases either. For the third time, but this time, with a grain of salt: it’s a demo.

DLSS 5’s technology isn't the death of artistic intent or the soul of the game. It, almost most certainly, will not cause all games to look the same. Not a catch-all-“everything goes through the exact same filter and all cartoon characters turn into hyper-realistic, uncanny, perfect facial features, millennial-looking white men” with how AI has historically been trained. DLSS 4 has essentially hit a peak with what it can do. DLSS 5 resets the scope and starts from the ground again. The wolf hasn't blown down the straw and stick house of DLSS 4. To future-proof, riding the wave of the rapid trends in AI development, Nvidia has already started building one with bricks. We just have to wait for the mortar to set.