NVIDIA’s DLSS 5 leak raises questions over AI-generated gaming visuals

NVIDIA’s upcoming DLSS 5 technology has appeared online ahead of its official release, giving players an early look at its approach to neural rendering. The leaked software has been tested across several games, but some results have raised concerns about visual quality, performance and the growing use of AI in game graphics.

The software was reportedly extracted from an early-access PC version of NBA 2K27 and subsequently adapted for use with other games. Because the build is unfinished and unofficial, the results may not represent the final version of DLSS 5. Nevertheless, footage and images shared online have highlighted some unusual changes to character faces and overall presentation.

DLSS 5 takes AI upscaling in a different direction

DLSS has traditionally focused on using AI to reconstruct higher-quality images from lower-resolution frames, helping games deliver improved image quality while maintaining higher performance. With DLSS 5, NVIDIA appears to be taking a more ambitious approach by using neural rendering to alter the appearance of scenes rather than simply improving the resolution of existing game data.

According to NVIDIA, the technology “infuses pixels with photoreal lighting and materials”. The aim is to make game scenes appear more realistic by using AI to add lighting and material details that may not have been present in the original rendering.

However, early examples suggest that the effect can sometimes move beyond realism and into something considerably stranger. Rather than making characters appear naturally lifelike, some results have produced faces that look as though they have been digitally transformed without corresponding changes to the rest of the character.

The effect can resemble an AI-generated interpretation of an existing game image. Character faces may become more detailed and realistic, but the surrounding models, animations and environments remain largely unchanged. This creates a visual contrast that can make the characters look less convincing rather than more believable.

Because the leaked software is not an official consumer release, these results should not be treated as representative of the finished technology. NVIDIA may make significant changes before DLSS 5 reaches players later this year.

Character faces highlight the uncanny valley problem

One of the more noticeable examples comes from Control, where the character Jesse Faden receives a substantial visual change when the leaked neural rendering technology is applied. Faden is intended to be around 28 years old in the game, but the AI-generated result can make her appear considerably older.

The transformation illustrates one of the challenges associated with using AI to modify existing character models. The system appears to recognise features in the original face and reinterpret them. Still, the result can shift the character’s perceived age and appearance in ways the game’s developers did not intend.

Skyrim provides another striking example. The game’s relatively old character models can look particularly unusual when combined with AI-generated facial detail. Characters can end up resembling realistic human faces placed over older, low-detail game models.

The problem becomes more obvious during movement. While the face may look more realistic in a still image, the underlying character model and animation system remain unchanged. Stiff movements, fixed expressions and unnatural eye behaviour can therefore clash with the newly realistic facial appearance.

That mismatch is at the centre of the uncanny valley effect seen in some of the leaked demonstrations. The technology can make individual elements appear more lifelike without making the entire character behave or move like a real person. Instead of producing a seamless improvement, the difference between the face and the rest of the model becomes more noticeable.

The approach could still have legitimate uses if developers integrate it carefully. A game designed around the technology from the beginning could potentially avoid some of the visual inconsistencies seen when the leaked build is added to older titles.

Early performance results add another concern

Visual quality is not the only issue raised by the leaked implementation. Performance figures circulating alongside demonstrations suggest that enabling neural rendering can substantially impact frame rates, although those results come from unofficial software and should be treated cautiously.

In one reported test, Control running on an RTX 5070 Ti at 4K allegedly fell from 71 frames per second to 35 frames per second after Neural Rendering was enabled. If accurate, that represents a significant performance reduction for a feature intended to enhance image quality.

Performance will be particularly important because DLSS has historically been associated with helping players achieve higher frame rates. Any new visual processing technique that consumes a large amount of GPU resources will need to provide a convincing improvement to justify the additional cost.

The final version could perform differently from the leaked build. Early software is often poorly optimised, and NVIDIA has not indicated that these unofficial tests reflect the performance players should expect from the finished release.

DLSS 5 is scheduled to launch officially in the autumn. Its eventual reception will likely depend heavily on how well NVIDIA balances its push toward AI-generated visual improvements with the qualities players value most, including consistent character design, natural animation, and strong performance.

Once developers begin implementing approved versions of the technology, it should become clearer whether DLSS 5 can deliver convincing improvements when it is properly integrated into a game. The leaked demonstrations may ultimately prove to be little more than an early experiment. Still, they have already highlighted the risks of using AI to make games look more realistic without changing the systems underneath.