I’ve tried to give DLSS 5 the benefit of the doubt, because on paper, it offers better performance, cleaner images, improved immersion at little cost (aside from needing NVIDIA’s most expensive GPUs), and that’s been the pitch for every version of DLSS so far, and broadly, it’s always worked to benefit gamers and developers.
But the more I’ve looked at the early DLSS 5 demos from GTC and how it’s currently being presented, the harder it is to ignore a more basic problem: it’s enhancing, yes, but it can also start to reinterpret the ‘human’ art. The changes feel disconnected from the tone of the scene, the character, and even the narrative context. In environments, DLSS 5’s use feels right, such as better lighting, more texture, improved depth and realism, but with characters, it feels wrong. Even major developers have been shocked by NVIDIA’s announcement.
Article continues below
After all, this Grace demo was the moment in the demos that stuck with me, and many others, because the changes are at odds with the scene’s intent: a character who’s clearly in the middle of something grim, who’s meant to look and feel distressed, is then reconstructed as with pristine, almost glamorous look, the kind of stylised aesthetic we’ve all started to associate with AI-generated portraits.
This is the problem with DLSS 5, as it has been shown, because that look isn’t coming solely from the developer’s assets but from how the model reconstructs them. It would appear that DLSS 5 isn’t rebuilding pixels from the game but instead leaning on patterns it has been trained to recognise, and those patterns are optimised for general image clarity and stability. The result is that familiar uncanny ‘AI’ aesthetic where everything veers toward the same idea of clean, attractive, well-lit imagery, regardless of whether it fits.
It’s hard to unsee, and once you clock it, it raises a bigger question about authorship and artistic intent, because if the final image is being shaped by data that exists outside the game’s art direction, then whose vision are you actually looking at? The developer’s or the AI model’s approximation of what it thinks looks right?
There’s always been a tension with these kinds of techniques, but DLSS 5 pushes it further. It’s not just stabilising or smoothing, it’s using AI-driven reconstruction that can influence the final image in ways that feel like artistic decisions. Which means there’s a deception to it that feels uneasy. And this is key, unlike some in this debate, I’m not staunchly against AI – I’ve seen how it can be used creatively, for example, the work of Kavan Cardoza – but there needs to be intent, openess and human creativity at the heart of its use.
NVIDIA, quite reasonably, says that developers can tune this, dial it back, and control its behavior, and I don’t doubt that’s true, to a point, but the demos don’t feel like isolated mistakes; they feel like a natural outcome of how the technology prioritises clarity. It’s doing what it’s designed to do, produce an image that looks ‘better’ according to its training and optimisation goals. The issue is that ‘better’ doesn’t always mean ‘correct’ and, in fact, when it comes to art direction, it often means the opposite.
A character’s face isn’t just geometry and texture detail, but rather it displays storytelling and context. A character’s design reflects where they’ve been, what they’ve gone through, and the tone of the world they exist in, and if a model comes in at the final step and subtly reinterprets that into something more generically appealing, then it’s diluting artistic intent.
That’s what makes DLSS 5 feel different from previous upscaling technology, as this isn’t like adding motion blur or smoothing textures, where the effect is clearly part of the original design. This is a layer that sits on top, reshaping the result based on learned reconstruction rather than purely the game’s authored pixels and styles.
If this becomes standard, if this is just how images are reconstructed going forward, then a layer of every game’s visual identity is effectively being filtered through shared reconstruction biases inherent to the model, and that means it’s not the developer’s choices, not the art team’s intent, but the model’s idea of what a good image looks like.
To be fair, there are still upsides to where DLSS 5 is going. The performance gains matter, especially as games continue to push hardware harder and costs rise, and in scenes where the reconstruction aligns well with the source material, the results can look genuinely impressive, as we saw in the Assassin’s Creed Shadows environment demo. But those wins don’t cancel out the underlying issue: it doesn’t just introduce artefacts or noise, it introduces the wrong idea and overrides the art direction a developer has curated, resulting in the wrong face or a mismatched tone.
It’s why the concern around DLSS 5 isn’t just knee-jerk anxiety about AI creeping into games and taking human jobs (though there’s a little of that) but a recognition that something fundamental is shifting, from rendering what was made to reconstructing what the system predicts should have been there, according to a system that wasn’t part of the original creative process.
I don’t think DLSS 5 is going away, and, in fact, I can see some developers designing games with the technology in mind, so those mismatches in tone and art direction aren’t an issue. It’s easy to imagine companies like Sony and Microsoft exploring similar approaches on it for next-gen consoles as well, in some form or other. I can also imagine NVIDIA creating deeper control options to ease anxiety and give game artists more input. If the control issue can be solved, it may not be as bad as we all think, but right now, that feels like a big ‘if’.
Visit the NVIDIA Blog for a full breakdown of the DLSS 5 tech, and make up your own mind.
