Frame by Frame: NVIDIA’s DLSS 5 and the Hidden Infringement Problem in Generative AI

In March 2026, GPU giant NVIDIA held its annual GPU Technology Conference (GTC) in San Jose, California. Capcom’s Resident Evil Requiem spanned the presentation stage screen. Deep Learning Super Sampling 5 (DLSS 5), NVIDIA’s new toy, had given the game’s central character a new look. Capcom’s preeminent protagonist Grace Ashcroft’s typically understated look was washed over with a hyper-realistic sheen: full lips, emotive eyes, textured hair with eyes pointing in different directions, and misshapen ears [1]. Animator Mike York called it “a complete AI re-render,” and “an AI filter over every frame” [2]. Earlier versions of DLSS upscaled resolution and interpolated frames from existing pixel data, reconstructing rather than creating. DLSS 5 is architecturally different: it takes a game's color data, depth buffers, and motion vectors, then uses a neural model with what NVIDIA calls “semantic understanding” to generate new visual information. Photoreal lighting, materials, and facial features are produced that never existed in the source art [3]. NVIDIA’s DLSS 5 exists at the intersection of three contentious areas of intellectual property (IP) law: derivative works, right of publicity, and AI training data. Current IP law is not equipped to address what happens when a third-party AI technology rewrites copyrighted content in real time, opening an enforcement gap with new dynamic elements, including frame-by-frame generation, software as distribution, and an output experienced only by the end user.

The legal problems DLSS 5 creates stem directly from a technical shift: the software is now generating image content rather than reconstructing it. The first version of DLSS launched in 2018 as an AI graphics upscaling tool. It was essentially an upgrade to a game’s existing visual data. The tool took low-resolution frames and used AI to reconstruct higher-resolution outputs to interpolate new frames between existing ones to smooth out performance [4]. The technology was a beautifier, refining preexisting content rather than generating anything new. NVIDIA founder Jensen Huang remarked that DLSS 5 is “the GPT moment for graphics,” telling GTC that NVIDIA has “fused controllable 3D graphics, the ground truth of virtual worlds, the structured data…with generative AI, probabilistic computing” [5]. NVIDIA demonstrated DLSS 5 running on titles from Capcom, EA, and Ubisoft, featuring endorsements from Bethesda’s Todd Howard and Capcom’s Jun Takeuchi [6]. However, art teams at both Capcom and Ubisoft were not informed before the conference demonstration. A Ubisoft developer explained that “we found out at the same time as the public” [7]. Similarly, Capcom’s art team felt “particularly shocked and worried” since the studios stated their anti-AI stance on their Resident Evil titles [8]. Hours after DLSS 5’s reveal, Jensen dismissed backlash at a press conference “Well, first of all, they’re completely wrong,” Jensen said. “It's not post-processing at the frame level, it’s generative control at the geometry level” [9]. His characterization was not merely imprecise. An NVIDIA engineer later confirmed that the model takes a single rendered frame as input and enters materials from that image alone, going beyond the “geometry level” inference Jensen claimed [10]. DLSS 5 does not read the game’s 3D models or material properties. Rather, it looks at the final picture, the flat image on screen, and makes inferences based on that. This distinction matters greatly; the technology is not enhancing the original asset, it is generating a new image based on it.

The U.S. Copyright Act of 1976 defines a “derivative work” as one “based upon one or more preexisting works” that is “recast, transformed, or adapted” [11]. The copyright holder possesses an exclusive right to create and distribute derivative works if they so choose [12]. For DLSS 5, the question is whether its outputs qualify as derivative or not. The affirmative case is fairly straightforward. DLSS 5 uses copyrighted game frames and transforms them, generating new facial features on copyrighted characters, altering lighting, and creating different material textures. The output is “based upon” the source artwork but contains substantial AI-generated additions that do not exist in the source material. In Andersen v. Stability AI (2023), a group of visual artists sued an AI company, alleging that its image-generation model was trained on their copyrighted works and produced infringing inputs. The Northern District Court of California allowed these claims to proceed, finding it plausible that AI-generated images based on copyrighted training data could constitute infringing copies or derivatives [13]. Fair use analysis under the Copyright Act asks whether the secondary use is “transformative,” meaning it adds something new with a different purpose or character as opposed to substituting content from the original. Additionally, under the Supreme Court’s recent decision in Andy Warhol Foundation v. Goldsmith (2023), transformative use requires more than simply adding new expressions when the secondary work serves substantially the same commercial purpose as the original [14]. Like the original game frame, a DLSS 5-produced output serves an identical purpose, displaying the game to the player.

The counterargument is equally as straightforward. The copyright holder, in this case the game publisher, may authorize derivative works. If Bethesda consents to DLSS 5’s modification of Starfield, no infringement of Bethesda’s copyright occurs. Under the work-for-hire doctrine, when a piece of work is created by an employee within the scope of employment, the employer is considered the legal author and owns all rights in the copyright [15]. Publishers typically own all game art created by employees, so an individual artist’s consent is legally irrelevant to the copyright. Importantly, however, games are not single-owner works. They contain a range of assets with a range of ownership chains such as licensed character, athlete likenesses, third-party music, middleware engines. A publisher can consent to modification of its own copyrighted material, but it cannot consent on the behalf of a licensor whose agreement did not consider AI alteration or extension. When Capcom endorsed DLSS 5 for Resident Evil Requiem, that endorsement only covered Capcom’s assets. The consent did not—and vitally could not—extend to every licensed element embedded in the game. The consent question demands an answer to whether a single authorization can cover the fragmented web of rights that a modern game contains.

The challenges associated with the derivative works problem are intensified when the licensed material includes real people. DLSS 5 altered the facial features of licensed individuals, including soccer player Virgil van Dijk in EA Sports FC, whose AI-modified face in the game was described by critics as that of “a different person” [16]. The right of publicity consists of a patchwork of state statutory rights protecting a person’s identity including name, image, voice, and likeness from unauthorized commercial use [17]. In Lehrman v. Lovo (2025), the Southern District Court of New York determined that right-of-publicity objection claims based on AI-generated voice clones could proceed, reasoning the New York’s statutory protections for name, voice, and likeness are “broad enough to encompass new technologies,” including generative AI [18]. For the fate of DLSS 5, if AI-generated voice clones implicate the right of publicity, AI-altered facial features would too. The problem is the consent chain. For example, imagine an athlete who licenses their likeness to EA, like Virgil Van Dijk. EA then participates in NVIDIA’s GLC demonstration. DLSS 5 modifies the athlete's likeness. EA’s specific likeness agreements are not publicly available, but standard industry practice predates generative AI and contemplates accurate depiction, not algorithmic alteration. On that assumption, the consent chain breaks at the moment DLSS 5 reshapes the licensed face. A likeness license grants rights for a specific, accurate depiction; an AI system that alters the licensed face into something the licensor could not have approved exceeds the scope of the original license. In the same way the Lehrman voice actors did not consent to having their recordings used to generate speech they never performed, an athlete who licenses their face for realistic depiction has not consented to having it reshaped. DLSS 5 tests this claim even more in that the alteration happens in real time, on every frame, without the licensor ever seeing the output. Potential federal legislation is seeking to resolve the uncertainty surrounding how existing likeness protections apply to AI-generated modifications. The No AI FRAUD Act and the NO FAKES Act would create federal causes of action for unauthorized AI-generated replicas of individuals, replacing the state-by-state patchwork currently in place [19] [20]. Both bills are still pending in Congress. The NO FAKES Act has not yet reached a floor vote after its reintroduction in April 2025 in both chambers. The NO AI FRAUD Act has stalled after being introduced in January 2024 and has not been formally reintroduced in the current Congress, though related bills addressing AI-enabled fraud continue to circulate. If enacted, such legislation would force DLSS 5 alterations of licensed likenesses to face federal liability, a consequence NVIDIA has not yet formally addressed.

The training data for DLSS 5’s neural model is a mystery outside of NVIDIA. The model possesses an understanding of faces, materials, and environments sophisticated enough to generate new visual content from a single frame, indicating training on innovatively large-scale visual datasets [21]. Meanwhile, the legal appetite for undisclosed AI training data is shrinking. In Thomson v. Ross Intelligence (2025), the District Court of Delaware granted summary judgment, finding that Ross’s use of copyrighted headnotes to train its AI was commercial, not transformative, and thus harmed the potential market for AI training data [22]. In Bartz v. Anthropic (2025), Judge William Alsup of the Northern District of California held that training on lawfully purchased books was “quintessentially transformative” and fair use, but training on pirated copies from shadow libraries was explicitly not, subjecting Anthropic to a $1.5 billion settlement [23]. The U.S. Copyright Office offered further explanation in its Part 3 Report, the third installment of its study of copyright and artificial intelligence released in May 2025. The office determined that multiple acts in the AI development pipeline, namely data collection, curation, and training, “can constitute prima facie infringement, particularly of the right of reproduction," meaning that each of these steps can satisfy elements of copyright infringement [24]. Because of the high likelihood that NVIDIA trained its rendering model on copyrighted game footage, film, and photography without authorization, the training itself is extremely legally vulnerable. This would particularly be where DLSS 5’s output directly modifies the visual content from which training data may have originated. The distinction from Bartz, between legitimately acquired training data which courts found could qualify as fair use, and pirated works which could not, is especially relevant here. Where the lawful acquisition of training data may survive fair use analysis, undisclosed acquisition almost certainly will not. NVIDIA’s insistence on not talking about their training data dilemma is not a legal defense, rather it highlights the liability they may soon face.

The controversy surrounding DLSS 5 surpasses complaints about its questionable aesthetics. DLSS 5 calls for a legal test on three of the most contested questions in AI intellectual property law: when AI modification of copyrighted material creates an unauthorized derivative, when AI alteration of a licensed likeness infringes the right of publicity, and whether fair use applies when AI output directly competes with the copyrighted original. Each question is unsettled, and DLSS 5 forces all three at once. With DLSS 5 scheduled for an official launch in Fall 2026, these questions will soon escape their academic boundary. The first lawsuit involving a licensed likeness altered using neural rendering may be the case that forces courts to finally draw lines that the current framework has left undrawn. The technology Jensen Huang called “the GPT moment for graphics” may also become a defining moment for generative AI intellectual property law.