DLSS 5 — When AI Seizes the Brush

The Ghost Author, Standardised Beauty, and the Twilight of Art Direction

This is the English translation of the original article written in French; the original version can be accessed at the following link.

[embed]DLSS 5 — Quand l’IA s’empare du pinceau L’auteur fantôme, la beauté normée et le crépuscule de la direction artistiquemedium.com

On March 16, 2026, at Nvidia’s GTC conference, the industry tipped into what might be called the era of post-human rendering.¹ If the demonstration covered many domains, one sequence crystallised my entire attention: the appearance of an EA Sports FC 26* athlete enhanced by this new iteration of DLSS 5. It is this image, asserting itself with the force of a technological mirage, that left the deepest impression on me. For someone like me, who is not a football connoisseur and does not know the precise physiognomy of current players, the sleight of hand is total. I saw in it the fulfilment of a fantasy native to simulation: an image so dense in organic detail that it seems to have reached its absolute horizon — the point where the virtual ceases to imitate the real and instead claims to replace it. The visual gap with the original game engine is abyssal.

[embed]DLSS 5 Announcement Video, Excerpt from GTC, March 16, 2026, @NVIDIAGeForce

Yet this apparent triumph conceals a far more unsettling mutation in the very nature of what we are looking at. This sense of stupefaction among connoisseurs arises from a precise rupture: where the layperson sees a feat, the expert denounces a profound alteration of identity and likeness. The AI no longer renders the subject — it reinterprets it according to the secret statistical references of its own training models, often in defiance of the original modelling. In pursuing absolute hyperrealism, the technology ends up betraying the very object it claims to magnify.

This paradox is at the heart of what follows. If DLSS 5 appears as an opportunity for small studios to reach otherwise inaccessible visual standards, it simultaneously opens a Pandora’s box for the industry’s giants, tempted to cut production costs by sacrificing creative singularity on the altar of universal smoothing. The question is no longer merely technical — it becomes political: must the image remain faithful to its author’s intention, or must it dissolve into a frictionless homogenisation, transforming video games into a succession of simulacra in service of a garish realism?

A clarification is necessary before entering the article’s development. This text does not claim that neural rendering is inherently condemned to normalised photo-realism. An AI model is not an aesthetic inevitability: it could, in theory, be trained on any visual direction. What I am examining here is the precise choice made on March 16, 2026 — its orientations and its consequences on the creative chain. Not what an ideal neural renderer might become in some speculative future, but what it does, now, in the form Nvidia has actually given it.

The Ghost Pixel: Anatomy of a Speculative Render

It all began as a technological crutch. Introduced in 2018 with Nvidia’s RTX 20xx series, DLSS — Deep Learning Super Sampling — had a pragmatic ambition: to allow games to run at high resolutions without mobilising the GPU’s full raw computational capacity. A simple upscaling technique at first, before Frame Generation began injecting ghost frames into the video stream. This process — generating two, four, or six artificial frames for each frame actually produced by the engine — opened the door to total substitution. With DLSS 5, the critical threshold is crossed: it is no longer just the framerate that is invented, but the entirety of the render. Nvidia cites the figure of 96% of pixels generated by the neural model rather than calculated by the GPU. This percentage is first and foremost a marketing argument, and it should not be made the centre of the reasoning. What matters is not the figure itself, but what it signals: an architectural shift in which the final appearance of the image is no longer produced by the GPU’s deterministic calculation, but by the probabilistic speculation of a model trained on corpora external to the game. The authentic image produced by the engine fades away; what replaces it raises the question of the sovereignty of the original work.

[embed]Explanatory video on DLSS in general and its version 5, @DrJVTek

To understand what this means concretely, one must enter the technical backstage: the G-Buffer, or Geometry Buffer. Before displaying a final image, the game engine generates a series of information layers that describe the scene in purely mathematical terms. This is not yet a visual image, but a precise mapping of the scene. It contains placement data: depth, which locates each object in space; surface orientation; normals, which indicate the inclination of each face; and motion vectors, which predict where each point will be in the next frame. This skeleton is completed by the physical properties of materials — a kind of DNA for surfaces. Albedo defines the raw colour of objects without any shadow; Roughness determines whether a surface is matte or glossy; Metallicity tells the light whether to bounce off a metal or be absorbed; and Ambient Occlusion maps zones of natural shadow in the creases and recesses of the geometry.²³

In Nvidia’s official discourse, DLSS 5 is described as a generative control system at the geometry level. In this ideal scenario, the AI would not merely skim over the scene; it would assimilate the entirety of the G-Buffer data and integrate it directly into the core of the final render generation. To use an organic metaphor: the game engine would provide the skeleton, and the artificial intelligence would fuse the physical properties to generate the muscles and flesh. The image would then be no longer a simple projection, but a reconstruction arising from a synthesis that crosses mathematical rigour with visual plausibility.

Yet the technical reality brought to light by the investigations of YouTuber Daniel Owen5 contradicts this version of a deep reconstruction. Owen revealed that DLSS 5 does not process geometry and material data directly, but instead receives as input a 2D image already rendered, accompanied by simple motion vectors. Since this description of the pipeline has to date received no denial from Nvidia, my analysis rests upon it. And what this nuance reveals is decisive: we are no longer in the realm of creating flesh or muscle, but in the application of a cosmetic skin — a sophisticated neural makeup. Instead of building the body from physical properties, the AI merely projects a hyperrealist mask over a simplified outline of the scene, analogous to pre-renders in animation. This compromise is dictated by a power constraint: processing all of this data in real time for every pixel would be too heavy a burden for current hardware. By limiting itself to reworking a 2D image, Nvidia saves resources at the cost of structural truth.

[embed]DLSS 5 Investigation Video, @danielowentech

It is this makeup — applied without genuine understanding of the underlying anatomy — that explains the visual drift documented during the first demonstrations, even if their attribution remains contested. The most emblematic example is that of a Starfield NPC, analysed by Daniel Owen⁵, whose haircut and temple structure were altered relative to the original model: where the native render showed a clean cut, DLSS 5 had modified its shape and appearance. Another video by the same analyst6 reveals an even more striking anomaly: a parasitic nasal protrusion — a ghost nostril that did not exist in the original model — emerging where the algorithm simply did not understand what it was looking at. Nvidia and certain analysts attributed these anomalies to the engine, Bethesda's Creation Engine, rather than to neural processing. But that is precisely where the problem lies: when almost the entirety of the image is produced by an opaque processing chain, it becomes impossible to identify who is responsible for what we see. The black box absorbs the fault as much as the authorship. One can no longer tell who betrays the image: the engine or the algorithm. This blurring of responsibility is, in itself, a confession.

This process resonates, for me — someone who works daily with ComfyUI* and image generation tools — with a striking familiarity. What Nvidia presents as an innovation is in reality the industrialisation of a logic that digital creators already practice: ControlNet conditioning. In these professional creative environments, image conception no longer rests on a simple text prompt, but on a system of nodes — small interconnected function boxes — that allow every stage of processing to be tuned with surgical precision. The artist provides the AI with a rigid structure, a kind of architectural blueprint that the generation must respect. This blueprint can take several forms: a depth map indicating what is near or far in the scene; an edge map that draws the silhouettes of objects; or a pose map, via a tool such as OpenPose, which precisely positions a character in space and dictates the position of each limb. Imagine a colouring book whose black lines represent not only shapes, but also depth and posture: the AI fills in the spaces while remaining strictly constrained by that framework. It acts like a virtuoso painter who injects colour and texture, yet remains strictly bound by the lines imposed upstream by the GPU — and does all this in 16 milliseconds, the time it takes the human eye to perceive an image, whereas an artist would have counted in hours.

[embed]ControlNet Tutorial for ComfyUI: Master Precise AI Generation, @aiandpixels

This process is in direct lineage with the traditional organisation of work in 3D animation and cinema VFX. For decades, the classic pipeline has required passing through a pre-render — a grey, bare version of the scene, without light or texture — to validate the staging, framing, and shot sequences. Only once this structure is approved do specialised departments intervene to embellish it: lighting designers, shading artists, compositors, each adding their layer up to the final image. The emergence of Video-to-Video* — an AI video generation technique whose source can be either a reference video that the model transforms and stylises, or a pre-render that the AI refines to deliver a fully finished version — opens this pre-render logic to a new domain: cinema, for the creation of sequences entirely generated by AI. DLSS 5 transposes this paradigm to video games in real time: the graphics card produces the instant previz, and the AI dresses it up to deliver the final enriched image.

It is precisely this logic that Nvidia attempted to promote during its GTC demonstration for Resident Evil: Requiem. By applying a systematic “beautification” effect to the heroine Grace Ashcroft, the algorithm smoothed her features to the point of giving her the look of a Snapchat filter, where the original render showed a face marked by exhaustion and emotional distress.⁷ Faced with the media firestorm, Nvidia presented an SDK — a software development kit offering studios control parameters over the neural pipeline: intensity adjustment, colour, contrast, saturation, and the ability to mask zones to be excluded from processing.8 For anyone working with tools like ComfyUI, these controls are immediately recognisable. The selection mask is nothing other than inpainting — the technique of protecting certain zones from model intervention. The intensity setting corresponds to denoise strength, a fundamental slider in any image generation pipeline. What is presented as innovation is merely the rudiments of AI generation, mastered for years by anyone who works seriously on these platforms. This corrective measure above all reveals an irreducible dilemma: if the AI must efface itself to preserve the author’s intention, the performance gain evaporates. If the AI is allowed to operate unconstrained, it ignores the features of the original design and substitutes its own norm of the real. The authentic image is reduced to a negligible residue.

Fundamentally, what this analysis reveals is that DLSS 5 reinvents nothing: it industrialises techniques already at work in AI image and video creation. The genuine feat — and it must be acknowledged — is execution speed. Where this process typically requires several seconds or even minutes, DLSS 5 accomplishes it in 16 milliseconds, the duration of a single frame. It is this transition to real time that constitutes the authentic technological leap. But behind this performance looms a more general and deeper question. If this finishing work — which once mobilised entire departments — can be accomplished in a few seconds by a generative model, what is the marginal value of those artists in the production chain? Human work undoubtedly retains a quality and depth that the machine has not yet learned to reach. But for mid-tier productions, the pressure is already real. And as the technology advances, the question will no longer be posed in merely economic terms: it is the very existence of these professions that the logic of automated rendering threatens to render obsolete.

The Strategic Dead End: When AI Conceals the Exhaustion of Silicon

The ubiquity of AI in our GPUs is not merely a quest for photorealism; it is above all a strategic response to a crisis in the silicon industry. Nvidia, aware that the raw power of its chips is now hitting the physical limits of matter, has set about transforming its graphics cards into vectors of a locked software rent.

The economic model of DLSS, since version 3.0, rests on an increasingly aggressive artificial segmentation. Each new generation of cards comes with an exclusive technological layer, creating a sharp break between old and new buyers. This system allows Nvidia to dictate and shorten hardware renewal cycles by using software as a watertight barrier — transforming what should be a driver update into a forced purchasing argument.

Yet the technical reality is far from as partitioned as the official discourse claims. At the launch of DLSS 3, Nvidia asserted that Frame Generation required the optical flow accelerator exclusive to the Ada Lovelace architecture (RTX 40xx and 50xx). This justification was rapidly undermined by the modding community: using third-party compatibility layers or driver workarounds, many users proved that these features were operational on supposedly obsolete hardware. The block is not a hardware inevitability, but a deliberate choice of commercial segmentation. This strategy masks an uncomfortable truth: raw rasterisation* power is stagnating. To justify ever-higher prices, Nvidia must oversell its proprietary AI layer, instituting firmware-driven planned obsolescence.

This lock-in is not new. For decades, Nvidia has cultivated a strategy of exclusivity. One recalls technologies such as PhysX or HairWorks which, though innovative, were deliberately restricted to run only on the manufacturer’s hardware, denying users of competing cards these visual experiences. All DLSS versions, from 1.0 onwards, have followed this trajectory: closed technologies, impossible to audit or port to other architectures. DLSS 5 pushes this logic to its apex: currently compatible only with the RTX 5090, it produces a striking paradox. If its primary purpose is to relieve the GPU by substituting calculation with AI, why does it require the fastest and most expensive card on the market? Two hypotheses emerge: either the generative model has become so heavy that it demands maximum raw power simply to imagine the image, or we are facing a purely marketing manoeuvre — Nvidia perhaps laying the groundwork for a future 60xx series, using DLSS 5 as bait. Unlike AMD, which with its early FSR versions had long advocated an open philosophy compatible with multiple hardware architectures, that manufacturer itself eventually adopted the same generational lock-in logic with FSR 4, now reserved for the RX 90xx series alone. The hermetic ecosystem is no longer Nvidia’s exclusive domain: it is the deep tendency of an entire industry.

Ray Tracing* and its ultimate evolution, Path Tracing, has long been presented as the grail of the absolute reconstruction of the physical truth of light. But this quest now runs up against the wall of thermodynamics and the exhaustion of Moore’s Law. Increasing raw power to simulate every photon now requires electricity consumption exceeding 500 or 600 watts for high-end models, and produces heat that silicon struggles to dissipate. Nvidia found itself in a physical impasse: how to keep promising the future of rendering when chips are incapable of calculating the billions of rays needed for real-time Path Tracing?

Before the great leap to DLSS 5, the solution passed through a series of technological patches, of which Ray Reconstruction was the contradictory apotheosis. To maintain fluidity, the GPU sends only a derisory fraction of the necessary rays, constraining the raw image to an unusable visual noise — a digital snowstorm too unpleasant to display as is. Ray Reconstruction then intervenes not as an additional calculation tool, but as an intelligent camouflage that smooths, fills, and invents the missing information, thereby “denoising” an image too harsh to show. The claim was to reach the physical truth of light; the reality was merely to disguise its absence beneath a layer of neural varnish. This technology was only a respiratory support for a render engine running on empty: the first admission — discreet but legible — that pure calculation had reached its limits.

DLSS 5 completes this observation with a definitive abandonment. Where previous versions still attempted to preserve some semblance of a link with physical simulation, it crosses the threshold of pure invention. Ray Tracing no longer serves the game engine: it has become a mere vestige. The engine produces a summary geometric pre-render; the AI slaps over it a hyperrealist layer generated from whole cloth, disconnected from any optical rigour. We have left deterministic rendering — calculated mathematically by the GPU — to enter the era of speculative rendering: imagined and “dreamed” by the AI.

Consent Violated

Behind the aesthetic controversy lay a scandal of a graver nature still. For if the technical controversy reveals the impasse of a technology at war with itself, the controversy reveals something more fundamental: systematic contempt for those whose work served as a showcase.

The public reaction was immediate and of an unprecedented scale for a technology announcement. The official DLSS 5 presentation video on YouTube reached 1.7 million views in a few days, but with a reception ratio that has no precedent in the history of hardware announcements: barely 16% positive reactions.⁸ The comments, of a rare unanimity, converged around shared vocabulary: “AI slop”, “yassification”, “generic beauty”, “the end of visual identity”. A single critical comment accumulated more than 23,000 likes — more than the total of all positive reactions to the video.

The specialist press was no less vocal. Digital Foundry, the uncontested reference in graphical analysis on YouTube, published an enthusiastic video in the hours following the announcement that collected 795,000 views in 21 hours,⁹ but also, according to estimates from the Return YouTube Dislike extension, nearly 59,000 thumbs down versus 24,000 positive. The editorial team even received threats — a fact that deserves to be named for what it is: an attempt at intimidation against the right to critical analysis, inadmissible in any context whatsoever. This prompted Digital Foundry, two days later, on March 18, to publish a partial retraction video unambiguously titled “Why We Should Have Waited…”, in which its graphical expert Alex Battaglia publicly acknowledged the ethical concerns raised by the technology and its capacity to “mess with artistic vision”. This public backpedalling from a medium as measured and technical as Digital Foundry constitutes, in itself, an admission of the depth of the unease.¹⁰

Behind these figures lay a scandal of a graver nature: that of consent. Windows Central quickly revealed that the artistic teams of the studios whose games had been featured in the Nvidia demonstration had not been informed in advance of their participation.¹¹ At Capcom as at Ubisoft, entire teams discovered their own games transformed by DLSS 5 at the same time as the general public — at their screens or via social networks. The phrase that circulated in professional discussion threads is lapidary: “We found out at the same time as the public”.

Kotaku, which conducted an in-depth investigation with several internal sources under cover of anonymity, obtained testimony of rare virulence, summarised in the very title of their article: “What the f***, Nvidia?”.¹² The fracture was not between Nvidia and the studios as a whole, but within those studios themselves: between executive management, who had granted their commercial approval to Nvidia, and the artistic teams, who had not been consulted. Jun Takeuchi at Capcom and Todd Howard at Bethesda are said to have given the green light, stepping over the teams who had concretely spent years working on the visual direction of these games.

The situation grew more complex when Jensen Huang publicly responded to the criticism. In a pointed statement, Nvidia’s CEO declared that detractors “are just completely wrong”, that the technology respected the artistic intent of developers, and that the concerns expressed reflected a fundamental misunderstanding of how it works. He reiterated that DLSS 5 fuses the controllability of geometry and textures with generative AI, and that developers and artists would have total control over the results.¹³

What followed was, for Nvidia’s corporate culture, an unusually embarrassing moment: a Nvidia technical representative publicly contradicted, on social networks, the official version his own CEO had just given of the neural pipeline. Without naming Huang directly, the employee specified that certain claims regarding the systematic respect of artistic intent and the extent of the control offered to developers were inaccurate — or at least greatly overstated relative to the official public presentation. This confirmed Daniel Owen’s analysis: DLSS 5 operates on an already-rendered 2D image, not on the deep geometric data Huang had claimed. This episode of internal dissent — extremely rare in a company as vertically structured as Nvidia — completes the portrait of a technology presented in marketing discourse that significantly exceeds its actual capabilities.

The Death of Art Direction

Beyond the technical feat and economic cynicism, DLSS 5 raises a fundamental question about the very nature of the digital work. By entrusting the quasi-totality of the image surface to an artificial intelligence, we are not merely changing the way games are made: we are changing the nature of what we are looking at. The image is no longer an intention sculpted in code — it becomes the result of a negotiation between an emaciated game engine and a sovereign algorithm.

To understand the rupture, one must analyse the nature of DLSS 5’s training. Early versions of the technology were trained on video game images: the AI learned to predict what a scaled-up game pixel should look like, respecting its visual codes. The DLSS 5 programme is radically different. To achieve the promised hyperrealist look, the model is no longer fed polygons but photographic databases. The AI is no longer asked to look like a game — it is asked to transform the game into a photograph. This shift is a direct assault on the visual singularity of each work. By plastering photographic truth over a virtual world, Nvidia ignores a fundamental rule: the purpose of a video game is not photo-realism, but aesthetic coherence.

Video games have always known how to achieve a form of truth without passing through photography. Zelda: Breath of the Wild is deeply realistic in its handling of physics, wind and light — but through a “Ghibli” aesthetic that constitutes its singularity. Red Dead Redemption 2 owes its sense of truth to an atmospheric management and colourist approach drawn from genre cinema, not from mere real-world capture. What the first demonstrations of DLSS 5 revealed with documentary precision is the erasure of these choices: each studio constructs, over months of work, a specific colour palette, a grading, a LUT — those invisible chromatic layers that give Cyberpunk its neon-tinted reflections, Resident Evil its grain and atmospheric tension, Red Dead Redemption 2 its patina of an old Western. In the demos, this layer appears neutralised, crushed by the model, which converges toward what it considers visually “beautiful” according to its own photographic references, rather than respecting the render intended by the studio. This is not a calibration flaw correctable by additional sliders: it is a collision of paradigms. On one side, an artistic pipeline founded on deliberate human decisions; on the other, an optimiser trained on a photographic norm exterior to the work. The artist constructs a gaze; the algorithm replaces it with an average.

The tension between an author’s original intent and the transformations imposed by third parties is not new in video game history. Modding culture has always inhabited it: re-shading, retexturing, graphical updating — interventions found in many successful games, from emblematic titles such as Skyrim, The Witcher 3, or Cyberpunk 2077. These mods sometimes aim to bring the game closer to a stricter photographic realism, sometimes simply to give it a different touch from that intended by the studio. This is notably the case of TapePunk, applied to Cyberpunk 2077: by substituting the game’s neon aesthetic with a raw VHS image — made of glitches and analogue noise — this mod paradoxically produces a sense of truth that high definition cannot achieve, as if degradation, in its very imperfection, reaches toward a form of authenticity. Where Nvidia believes the real resides in an accumulation of statistical details, TapePunk* reminds us that the real is often lodged in what is absent. But in every case, the modder’s approach remains conscious, optional, transparent. DLSS 5 breaks radically with this logic. It does not present itself as an aesthetic choice but as an infrastructure. It is no longer the user who selects a filter: it is the manufacturer who decides, upstream, what the image will be allowed to show.

[embed]TapePunk Mod Demonstration, @gamegpu

This evolution redefines the role of technical artists as curators of hallucinations: the artist no longer creates the right image — they must choose the least-wrong one from among the generated proposals. Creative work is transformed into a form of geometric prompt engineering, where one no longer adjusts lights or materials but sets selection masks and denoising weights to try to force the AI not to stray too far from the initial vision. It is a security guard’s job: monitoring that the algorithm does not alter a haircut, that it does not transform a dank, narrative atmosphere into a cosmetics advertisement. One gains in statistical detail what one loses in evocative force. Ultimately, DLSS 5 sketches a future where the game engine uses increasingly minimalist structures. If the AI is capable of generating a hyperrealist image from a simple geometric skeleton, why burden oneself with heavy physics engines? One can imagine an extreme version of the video game where the graphics card receives nothing but summary data streams to generate a real-time high-fidelity video: the absolute triumph of image without engine, of work without identified creator.

Some see in this mutation not a death of art direction but the dawn of a new space of control: if tools improve, if AI learns to respect a studio’s LUTs and grading, might it not become a genuine lever in service of artists? The argument is not without foundation and it would be intellectually dishonest to ignore it. But it calls for a precise response. First, this ideal future does not yet exist, and it is the present that must be analysed, not a promise. Then — and this is essential — even in this optimistic scenario: who decides the parameters? Who controls the model? Who defines the acceptable margin between the intended image and the generated one? The answer today is the same as yesterday: Nvidia. Technology is not neutral, and promising artists future control tools does not change the fact that the generative layer will remain proprietary, opaque, and deployed in service of a manufacturer’s hardware segmentation. What DLSS 5 inaugurates is not a partnership. It is a dependency.

[embed]DLSS 5: NVIDIA will RUIN my Art Direction! @Gagzzz82

Roland Barthes, in The Death of the Author (1968),¹⁴ had theorised the disappearance of the creator as a liberation: that of the reader, finally free to produce their own meaning, to weave multiple significations from a text whose author could no longer close off interpretation. For Barthes, this death was productive, almost joyful. It opened the work to the world. DLSS 5 accomplishes this death, but in a radically different register. The disappearance of the artist there does not liberate the player: it installs in their place a proprietary algorithm that decides, alone, what the gaze must see. It is the advent of the machine as sovereign author — without intention, without biography, without responsibility, and above all without the possibility of being questioned. Barthes celebrated the polyphony of meaning; DLSS 5 imposes the statistical monophony of beauty.

In traditional cinema or the pre-generative video game, an author’s signature was lodged in the mastery of detail: the choice of a focal length, the temperature of a colour, the roughness of a texture. It was a form of sovereignty over matter. With DLSS 5, this sovereignty evaporates. When the quasi-totality of the image is produced by an AI trained on global photographic databases, the work becomes a forced collaboration with the machine and, ultimately, with Nvidia the multinational. The artist is no longer the one who makes — they are the one who validates. This transformation of the creator into a curator of hallucinations poses a major ethical problem: can one still speak of art direction when the essential beauty produced is a statistical reminiscence of a third party’s database?

Faced with this hegemony of the standardised render, true artistic independence will reside in affirmed stylisation. Resistance will pass through a radical design decision that renders DLSS 5 structurally obsolete. By choosing an aesthetic that deliberately distances itself from the canons of the real — abstraction, flat colours, assertive graphics — the creator breaks the very logic of the algorithm. Since DLSS 5 is trained on photographic databases to beautify the real, it becomes incapable of processing a work that refuses these codes. Applying a photo-realist “beautifier” to an expressionist work would only destroy its intention.

This resistance through form may be the most just response, but it remains to be constructed, invented, and collectively assumed. It does not say everything, either, about the future of the technology itself. The history of great hardware ruptures invites us not to reduce this analysis to the short term alone. Every major technological wave was first captured by large actors before democratising, miniaturising, and ending up in the hands of creators. It is possible that, in time, studios and art directors will reappropriate these tools, that more open models will emerge, that Nvidia’s closed logic is only an initial phase. This perspective deserves to be heard. But Nvidia’s track record argues poorly for a scenario of rapid opening: technologies locked generation after generation, studios kept uninformed, commercial segmentation erected into a system. It is on this short term — documented and observable — that the essential part of this article rests. The future will tell us whether neural rendering follows the trajectory of general computing toward democratisation, or that of the cultural industry toward concentration. I remain, for my part, pessimistic — and this pessimism, as one widens one’s gaze beyond video games toward creative practice in general, only deepens.

For this question extends far beyond the frame of video games. It touches something more fundamental in every creative practice: the transmission of gesture, learning through friction. What forms a gaze — whether in cinema, photography, visual arts, or design — is precisely the resistance of matter, the error that forces understanding, the imperfection that reveals an intention. When a tool promises to magnify everything at the press of a button, it is not only a technical know-how that fades: it is the awareness of what one is doing and why. The risk is not producing less skilled creators. It is producing creators who will no longer know how to distinguish their own intention from the machine’s proposal.

The Hegemony of the Gaze

The advent of DLSS 5 does not merely mark a technological milestone: it inaugurates what seems to me a new ontological regime for the image. What follows is a personal reading, drawing on the reflections of thinkers who analysed the logic of cultural industries long before the AI era. It is not the only possible reading. But it is the one that seems to me most coherent with the direction history takes when read from the field of art, culture, and capitalism. If one accepts the idea that the quasi-totality of what we see is no longer the fruit of deterministic calculation but of neural speculation, it is our contract of trust with the medium that crumbles. We pass from an indexical image — the pixel as proof of a mathematical datum — to a spectral image, a hallucination of what it should be.

Traditionally, the player believed in the image because they knew that behind each reflection lay a rendering equation. With DLSS 5, this link is broken. The AI no longer shows the world: it convinces the brain of its presence through probabilistic plausibility. This is the very definition of the simulacrum as Baudrillard theorised it¹⁵ : a realism that feeds on itself to ultimately occlude the world it claimed to represent. We no longer contemplate the game scene; we contemplate what Nvidia’s algorithm thinks a game scene should statistically look like. The map is no longer the territory. It has replaced it.

This transition opens a perspective more clearly perceptible from the field of cinema. If the game engine serves only to provide a geometric skeleton for real-time video synthesis, the video game progressively abandons its computational nature to become a flux — a liquid image, without fixed edges, flowing according to statistical probabilities rather than according to a constructed intention. The boundary with AI video creation tools such as Sora or Runway evaporates. The video game becomes a succession of on-the-fly generated shots, a staging without a set, an image without a camera. For the player, the experience is transformed: they no longer navigate a persistent and precise space governed by coherent physical rules, but in a flow of visual plausibilities. The risk is to see the artist’s trace disappear in favour of a liquid aesthetic capable of mimicking everything while never asserting anything solid — an image that adapts, flows, but refuses to solidify into intention.

This shift has immediate and brutal economic consequences. Electronic Arts devotes, for each edition of its football franchise, colossal resources to the scanning and modelling of players’ faces. Entire teams of character artists, shading artists, lighting artists spend months sculpting models, refining textures, calibrating materials so that a footballer’s skin looks like skin and not wax. The result, year after year, runs up against the same obstacle: the persistent uncanny valley that the gaming community documents with cruel regularity. In a few frames, DLSS 5 produces a result that objectively surpasses years of human work. The question that follows is economic as much as artistic: if AI accomplishes in post-processing what dozens of artists cannot achieve upstream, what is the marginal value of those artists in the production chain? This reasoning follows exactly the same trajectory as the introduction of generative AI tools in concept art production: one no longer sells a talent, one sells productivity enhanced by the tool, thereby justifying a drastic reduction in headcount.

Defenders of DLSS 5 often advance the democratisation argument: a team of five people could now obtain a render comparable to that of a studio of two hundred. But this promise runs up against artistic reality, one inscribed in a levelling toward photographic realism that is precisely the enemy of creative differentiation. The most striking independent successes of recent years — Hades, Hollow Knight, Disco Elysium, Outer Wilds — are not distinguished by their mimicry of the real, but by a singular and irreducible aesthetic vision. DLSS 5 does not democratise art. It democratises the norm.

It is here that Adorno and Horkheimer, in their Dialectic of Enlightenment,¹⁶ seem to be writing for today. They described the cultural industry as the large-scale production of standardised goods, reducing works to interchangeable variants of the same schema. DLSS 5, by imposing a layer of generic photographic realism over the whole of video game production, produces exactly this effect: an illusion of visual richness covering a growing uniformisation. When a sufficient number of games has been rendered through the same neural model, trained on the same photographic corpora, the visual diversity of the medium will be nothing but a memory. In wanting to make everything real, one ends up making everything identical.

This movement of homogenisation is not neutral. One must regard DLSS 5 as a political object. By imposing a standard of representation based on real photography, Nvidia is not merely exporting a technology: it is exporting a worldview, a gaze that privileges clarity over shadow, norm over margin, statistics over emotion. When a single company controls the algorithms that generate the quasi-totality of what we see in our interactive fictions, it exercises a power of normalisation without precedent: that of a cultural black box whose training corpora we do not know, whose aesthetic biases are opaque, but whose visual prejudices we suffer daily. The video game — once a space of virtually infinite plastic experimentation — risks closing in on a single photographic ideal, transforming the medium into a vast hall of mirrors where the AI returns to us only the average of what we have already seen.

By correcting a heroine’s features to bring them into line with an algorithmic ideal, by smoothing the distress of a face until it is indistinguishable from a social media filter, the AI distances us from the organic truth of the living. It projects us into an aseptic world where we no longer encounter the Other in their difference, but only the aggregated reflection of our own data. Bergson, in Creative Evolution,¹⁷ defined life as duration — an unpredictable welling-up, a continuous creation of new forms. Nvidia’s algorithm calculates the inverse: not the exception, but the consensus; not the unforeseen that defines the living, but the statistical majority that freezes it. This is not a render. It is an average.

The true urgency is no longer to gain frames per second, but to regain our visual sovereignty. DLSS 5 promises us a perfect world; our duty as creators is to recall that art begins where perfection ends.

Switching Off the Black Box to Rekindle the Gaze

The DLSS 5 affair cannot be reduced to a simple controversy over the performance of a high-end graphics card. It must be read as the symptom of a deeper metaphysical shift: that of the systematic delegation of the creative gesture to statistics. In crossing the threshold of the quasi-totality of generated pixels, we are no longer in the realm of technical assistance. We are in ontological replacement.

Until now, the GPU was a scrupulous tool — a mathematical slave in service of human intention. It calculated what we ordered it to calculate, nothing more. With generative AI integrated at the heart of rendering, this tool overflows its function. It no longer merely executes: it interprets, proposes, hallucinates. Can one still speak of cohabitation when the tool appropriates the quasi-totality of the work’s sensitive surface? The promise of collaboration between human and machine is a lure if the machine holds the monopoly on finishing and beauty. The brush has become the artist. The human, in turn, has become the flow manager — a black box supervisor dispossessed of the materiality of their own work.

What DLSS 5 reveals, at bottom, is our relationship to imperfection. AI, in its current nature, is intrinsically normative: it imposes its own paradigm of the average as universal truth. By feeding on billions of images to define what a face, a light, a texture is, it evacuates the accident, the fault, the singularity — everything that, precisely, constitutes the essence of the creative gesture. Through DLSS 5, it is our relationship to materiality and to proof that collapses. If the image is no longer the index of a reality — even virtual and calculated — but a pure speculation, then the link between seeing and knowing is broken. AI encloses us in a realism without reality. Art has never been an average. It has always been a deviation.

For the creator — filmmaker, visual artist, designer, photographer — the stakes now extend far beyond the frame of video games. It is a matter of defending a certain idea of what making means: accepting the unforeseen, the unplannable, the laborious. Preserving spaces of creation where AI remains in its place as a servile tool, never crossing the frontier of intention. This is not a technophobic struggle: it is an ethical imperative. That of refusing to confuse the machine’s proposal with an artistic decision.

Switching off the black box to rekindle the gaze. The challenge is not technical. It is political, aesthetic, and profoundly human. For if the image ends by doing without the world, the risk is that, in time, the world will end by doing without us — not in a brutal collapse, but in the silent erasure of everything that, in a work, resists the average.

Références

1. NVIDIA Newsroom, NVIDIA DLSS 5 Delivers AI-Powered Breakthrough in Visual Fidelity for Games

2. Tom’s Hardware, We got a first look at Nvidia’s DLSS 5 and the future of neural rendering at GTC

3. FXguide, NVIDIA’s new real-time neural rendering with DLSS 5

4. Nvidia Answers my DLSS 5 Questions, @DanielOwen, march 2026.

5. FSR 4.1 Drama, and Nvidia tries to explain DLSS5, @DanielOwen, march 2026.

6. PC Gamer, DLSS 5 clearly overwrites game characters with AI beauty standards, but Nvidia says devs have ‘artistic control’

7. 80.lv, NVIDIA Says DLSS 5 Was “Designed With Developers” & They Have Creative Control

8. VideoGamesChronicle, Nvidia’s DLSS 5 reveal trailer only has 16% likes on YouTube as players make their voices heard

9. @DigitalFoundry, 16 march 2026. Hands-On With DLSS 5: Our First Look At Nvidia’s Next-Gen Photo-Realistic Lighting

10. Why We Should Have Waited With Our Coverage, @DigitalFoundry, 18 march 2026 ; NotebookCheck, Digital Foundry’s Alex Battaglia calls DLSS 5 out for messing with artistic vision and serious ethical concerns

11. Windows Central, NVIDIA DLSS 5 reveal left devs in the dark

12. PC Gamer, ‘Bad ending: now every game is slop’: Game developers share mixed reactions to DLSS 5

13. HotHardware, NVIDIA DLSS 5 Backlash: Jensen Huang Says ‘AI Slop’ Critics Are Completely Wrong

14. Roland Barthes, The Death of the Author, Manteia, n° 5, 1968.

15. Jean Baudrillard, Simulacra and Simulation, Galilée, 1981.

16. Theodor W. Adorno & Max Horkheimer, Dialectic of Enlightenment, Gallimard, 1974 (original edition : Dialektik der Aufklärung, 1944).

17. Henri Bergson, L’Évolution créatrice, PUF, 1907.

Glossary of Technical Terms

Color grading: process of adjusting the colours of an image or video sequence to create a coherent visual atmosphere and recognisable aesthetic identity. Distinct from simple colour correction, it reflects deliberate artistic intent.

ComfyUI: open-source graphical interface allowing users to build AI image generation pipelines. Widely used by creators for inpainting, upscaling, and Video-to-Video operations.

Frame Generation: a DLSS feature that generates intermediate frames between two GPU-calculated frames in order to artificially increase the displayed frames per second.

GPU (Graphics Processing Unit): a graphics processor dedicated to real-time image calculation. Unlike the generalist CPU, it is designed to process millions of parallel operations simultaneously.

GTC (GPU Technology Conference): an annual conference organised by Nvidia, devoted to advances in artificial intelligence and graphical computing. DLSS 5 was presented at the March 2026 GTC.

LUT (Look-Up Table): a chromatic correspondence table applied in post-production to give a specific colourist identity to an image. Used in video games, cinema, and photography to create a coherent and recognisable visual identity.

Rasterisation: traditional 3D rendering technique that consists of projecting three-dimensional objects onto a 2D surface pixel by pixel. The basic method of video games since the 1990s, it is today complemented or replaced by neural approaches.

Ray Tracing: a rendering technique that simulates the real physical behaviour of light: reflections, refractions, cast shadows. Very precise, it is also very resource-intensive, which led Nvidia to develop dedicated hardware accelerators.

Re-shading: real-time modification of lighting, contrast, and visual effects in a game via third-party tools such as ReShade. Common practice in the modding community to personalise the visual appearance of a game.

RTX / Ada Lovelace Architecture: Nvidia’s range of graphics cards incorporating dedicated cores for Ray Tracing and AI (Tensor Cores). The Ada Lovelace architecture refers to the chip generation introduced with the RTX 40xx series in 2022, on which certain exclusive DLSS features depend.

Runway: an AI-assisted video creation platform enabling the generation and transformation of video sequences from images or textual descriptions.

Sora: an AI video generation tool developed by OpenAI, capable of producing realistic video sequences from textual descriptions or reference images.

Upscaling: technique of increasing the resolution of a low-resolution image to a higher resolution, either by classical interpolation or by machine learning in the case of DLSS.

Video-to-Video: an AI video generation technique in which the input source is an existing video. Depending on the use case, this source can be a reference video that the model transforms and stylises, or a pre-render that the AI refines to produce a finalised version.