The Case of the Stolen Artistic Vision
Ahead of the official release of DLSS 5, NVIDIA held another presentation of the technology. It once again stirred up public opinion on the Internet: whether such technology is needed or not, and how this technology will ruin games or make them better.
NVIDIA probably faced such a discussion on this scale for the first time. When DLSS 5 was first introduced to the public in March 2026, a scandal almost erupted online. Some developers and studio representatives even rushed to distance themselves from certain aspects of this presentation, stating that they were unaware of the upcoming demonstrations or saw them simultaneously with the public. So powerful was the wave of negativity from players who compared the demonstrated game footage to “AI slop.”
Jensen Huang even complained to the press in the same March that the company was misunderstood. Or rather, he spoke even more harshly about critics: “Well, first of all, they're completely wrong.” And the recent presentation is essentially an extended version of this thesis. In it, NVIDIA specialists try to actually prove that… But what exactly are they trying to prove? We will answer this question a little later. For now, let's ask another question.
Why does NVIDIA need to prove anything? If you don't like DLSS 5 technology, don't enable it. As the apologists of DLSS 5 say, the dog barks, but the caravan moves on. But for some reason, specifically in the case of DLSS 5, the company persistently tries to prove its usefulness. In the case of DLSS 1–4, there was no such need for such active defense of the technology's concept. What is the reason?
It's hard to argue with the fundamental usefulness of DLSS 1–4. Criticism of these technologies usually boiled down to specific implementations and problems with specific implementations. Namely: a somewhat blurry picture during upscaling (DLSS 1), latency issues when redrawing frames (DLSS 3), and some other rough edges. But overall, the community accepted this approach, and generally, the attitude towards DLSS and competing solutions like FSR or XeSS comes down to the phrase: “Why not?” Enabled and saw no problems? Left it on. Are there problems? Turned it off. That's all.
The second point: the picture with RTX enabled in some (we emphasize, in some) games is indeed prettier, and enabling DLSS allows playing with a comfortable frame rate even on a relatively inexpensive graphics card.
That is, DLSS 1–4 primarily acts as a tool to improve performance. What was previously achieved by increasing the number of operations that a chip can perform per clock cycle, and by increasing the frequency itself, can now to some extent be transferred to the software plane. We leave part of the complex work to the chip — rendering — and then restore the missing information using neural network algorithms. And starting with DLSS 3, we also generate intermediate frames taking into account object and camera movement.
But the more work we shift to the “software,” neural network part, the more hardware support this software part requires. And this is where it gets interesting. From generation to generation, it's not just the number of specialized blocks that changes, but the very ratio of GPU computing capabilities: tensor computations are becoming an increasingly powerful and increasingly significant part of the overall computing architecture.
And this, perhaps, is the key point. With each generation of NVIDIA cards, not only the capabilities of traditional rendering grow, but also the computing capabilities of specialized neural network blocks: new generations of tensor cores appear, new computation formats, and increasingly high peak performance in AI tasks. As a result, the architecture of 3D graphics cards has an increasing amount of specialized computing resources that need to be used beyond traditional rendering. And DLSS, from a tool that allows relatively cheaply to get more frames, gradually turns into one of the ways to utilize this resource.
Why does NVIDIA need this?
So NVIDIA finds itself in a situation where it needs to sell graphics cards that are becoming more expensive from generation to generation, while not demonstrating a fundamental change in graphics quality in traditional rendering. And here, it would seem, a solution has been found. DLSS 5 is capable of making graphics more realistic. The neural network changes the appearance of materials, lighting, and image details — and the picture can indeed be transformed. NVIDIA itself positions DLSS 5 precisely as a transition from ordinary image reconstruction to generative neural rendering, which should add in real-time those lighting and material features that traditional rendering is forced to simplify.
And in response — a massive wave of criticism. Moreover, the criticism turned out to be so strong that the dispute went far beyond the usual discussion of the quality of the next version of DLSS: disputes began about where image reconstruction ends and generation begins, how much the neural network preserves the artistic intent of the developers, and whether such processing is even needed in games. The authority of Digital Foundry, whose assessments of game graphics quality were previously perceived by many as a kind of standard, also came under fire.
To this, it should be added that today NVIDIA sells not just graphics cards to the market, but the idea of a large-scale transition to neural network computing. The company's main business is increasingly connected with AI infrastructure, and public and political resistance is indeed growing around the construction of data centers: in different regions, people and authorities are already arguing over electricity consumption, strain on power grids, and other consequences of the rapid growth of data centers.
And here another problem arises for NVIDIA. An ordinary person may not see any direct benefit from another giant data center consuming a huge amount of energy. Therefore, it is important for the company to show the advantages of AI not only to corporations that buy its accelerators but also to ordinary consumers. And video games here are an almost ideal demonstration platform. DLSS literally allows a graphics card owner to see: here the neural network is working, here it saves computations, and here the picture becomes better with its help. AI is wonderful.
And if, in response to this demonstration, the player, instead of saying “look what AI can do now,” says: “Why do I need this?”, then for NVIDIA this is already more unpleasant than a usual dispute about the quality of another technology. Because in this case, not only the specific DLSS 5 is questioned, but also the very idea of moving computations deeper and deeper from traditional rendering into the neural network plane.
Attempts to whitewash the technology
And now, let me return to the question of what NVIDIA specialists are actually trying to prove in the latest presentation.
They are trying to prove that it's worth it.
After all, what claims were made against NVIDIA after the first presentation?
- The neural network changes the geometry of the scene.
- The neural network will produce different results with different runs.
- The neural network takes away the developer's control over how the game should look.
- The neural network can actually replace the original artistic intent with its own idea of how the scene should look.
And throughout the video, NVIDIA specialists answer these questions.
- No, the neural network does not change the original geometry of the scene. NVIDIA shows that DLSS 5 should not arbitrarily change the geometry and structure of the original scene. The geometry remains as the artists created it, and DLSS 5 primarily changes the materials' reaction to light and adds appropriate visual details.
- No, the result of the neural network's work is not random. NVIDIA emphasizes that DLSS 5 works deterministically and is designed for image stability from frame to frame.
- No, the developer does not lose control over the result. DLSS 5 provides several models, which can be configured differently for different scenes, and the intensity of structural and tonal changes can be adjusted separately.
- No, the developer is not obliged to give the entire picture to the neural network. There is semantic and manual masking, allowing to determine which objects and areas of the scene to apply neural network enhancements to, and which to leave practically untouched.
- And finally, no, DLSS 5 is not a completely autonomous image generator that doesn't care what was in the original frame. NVIDIA specifically emphasizes that the result remains tied to the base image and game engine data. Moreover, the higher the quality of the source data — for example, if the game already uses ray tracing or path tracing — the more accurate the final result.
That is, the August presentation is actually structured as a response to the initial wave of criticism.
By the way, we already asked some of these questions in our March article, when we analyzed the first DLSS 5 presentation. And the new NVIDIA presentation indeed provided answers to some of them.
In particular, it became much clearer that DLSS 5 is not a “magic button” that can simply be pressed and forgotten. Developers will have to configure the neural network, select models, adjust the intensity of processing, and, if necessary, limit its application to individual objects and areas of the scene. In other words — work. What prevents working on optimizing traditional rendering? However, here, perhaps, we will agree with NVIDIA. This tool is more convenient. But users of weak graphics cards will be very disappointed.
It became clear that the overexposure and deep shadows in the March presentation were a flaw. NVIDIA's tools allow for finer lighting adjustments. Why wasn't this done in March? Wearing a tinfoil hat, one might assume that such tools didn't exist then and only appeared by August.
Taking off the tinfoil hat. NVIDIA now directly shows that the developer can control how much the neural network interferes with the image. We want to believe that in real games, this will indeed be used, and strange overexposures, shadow clipping, and other artifacts from the first demonstrations will be minimized.
But whether this will work exactly like that remains to be seen.
Because the problem with DLSS 5, as we see it, is not in the technical details and implementation of a specific technology. The problem is that players don't like the result of the neural network's work.
Is the player not the same?
Last time we also raised the question of the uncanny valley effect. When character images become too realistic, but animations remain as they are, the character on the screen can start to be off-putting. This is no longer a question of texture quality, lighting, or polygon count. The brain notices a discrepancy between the realistic appearance and the unnatural behavior of the character, and they begin to seem strange, wrong, sometimes even frightening. The more realistic the appearance becomes, the more noticeable the imperfection of the animation can be.
And this is not just our subjective impression. Players and journalists are increasingly talking about the uncanny valley in games with ultra-realistic graphics. What to do in this case? NVIDIA essentially did not answer in the latest presentation.
Because DLSS 5 can improve skin, eyes, hair, lighting, and reflections as much as it wants, but the technology itself does not fix character animation. If the character continues to move unnaturally, the neural network can only make their face more realistic — and thereby further emphasize the problem.
In addition, there is still one question we asked in March: how will DLSS 5 behave in dynamic scenes?
And here's a telling detail: even in the latest NVIDIA presentation, the main focus is again on comparing individual scenes and visual details. We are shown how much better the skin looks, how the lighting changes, how hair, materials, and reflections are rendered. But when it comes to the main thing — how all this looks in real dynamic gameplay — significantly more questions remain.
At such moments, the eye doesn't catch details at all. A player may be completely indifferent to how realistically the neural network drew the character's skin or the reflection of light on their clothes if they only get to see it for a fraction of a second.
And a rather amusing situation arises. NVIDIA has put enormous effort into proving to us that the neural network can make an image more realistic. But the main question may turn out to be completely different: will the player even notice this when a normal game begins? And if they notice the work of this technology in cutscenes, will they be happy about it?
Long before the DLSS 5 presentation, specifically in Maxim Ivanov's article, a more general question was raised: is realism in games, and graphics realism in particular, even necessary? Here I will allow myself to quote from that material:
Realism in games can manifest in any way. It can be objects obeying the laws of physics, characters behaving like your neighbors, the ability to turn on water, light, or see yourself in a mirror. These are tools whose skillful use helps developers immerse the player in what is happening. And in particularly skilled hands, “realism” becomes the main focus for a good game, without violating the main rule: games are fun!
But it is worth remembering that the very essence of games is to escape from reality. These are portals to other worlds, where the scale of what is happening is limited only by the developers' imagination. Realism can be both a support and a ballast for such worlds. And if it is used thoughtlessly, players will prefer simple, but fun and understandable games to super-realistic elements.
Perhaps this will sound bold, but it's a pity that NVIDIA's bosses don't understand this. And even worse, if they do understand it, but pretend that the criticism of DLSS 5 boils down to a simple set of technical claims that can be answered with another presentation.
Possible, but why?
Players, for the most part, are asking the wrong questions, and NVIDIA is giving the wrong answers.
NVIDIA has indeed answered most of the technical claims. Moreover, after the August presentation, it becomes much harder to argue that DLSS 5 is some uncontrolled neural network that takes an image as input and outputs God knows what. No. The developer has control tools, there are different models, there is the possibility to limit the scope of the technology. NVIDIA has clearly done a lot of work to address precisely these questions.
But we still haven't been explained why the player should want their game to look exactly like this. We were told in detail how to make the neural network make the image more realistic: that it can change skin, hair, materials, lighting, and reflections. And even the developer will be able to control this.
But the question “Why?” has not gone away. If NVIDIA thought that the new presentation would solve all problems and players would love “AI slop,” then it seems they failed. And this is a problem that cannot be solved by a new model, masking, or additional developer control.