8/31/26

DLSS 5 leaks before release: modders run NVIDIA's neural rendering in other people's games and lose half their frames

Interesting From The Web

NVIDIA's neural rendering DLSS 5 has spread across the web ahead of its official release. In the installation files of NBA 2K27, which has launched in early access, dataminers found the nvngx_dlssnr.dll library, and modders have already run it in games that nobody adapted for DLSS 5. Over the weekend comparisons from various projects appeared online; the first public example was Control.

For the first time the technology is being evaluated on arbitrary scenes from other people's games rather than on clips prepared by NVIDIA. Both sides are visible: DLSS 5 noticeably rebuilds lighting and character faces and at the same time cuts the frame rate roughly in half even on current-generation cards.

How the DLSS 5 runtime ended up in other people's games

The nvngx_dlssnr.dll file turned up shortly after the start of NBA 2K27's early access. Its size is 158 MB, the largest library among all released DLSS versions; according to analysis, most of that volume is taken up by the neural network's weights, and the build itself appears not to be final. There is no direct mention of DLSS 5 in the library, but there is the abbreviation NR — presumably neural rendering, NVIDIA's name for the technology that reconstructs lighting from the frame and motion vectors. There are no DLSS 5 options in the game itself, so it is incorrect to speak of the technology being supported in NBA 2K27.

The library was picked up by the developers of the RenoDX mod for Reshade, who quickly got it working in Control. The mod has a set of sliders and toggles — this confirms NVIDIA's earlier promise that developers will be able to change the intensity of the effect and exclude individual objects from processing.

Faces are again at the center of the argument

Videos with the mod show that even at relatively conservative settings DLSS 5 noticeably changes Control's lighting and the characters' faces: shadows are added and deepened, skin shading behaves differently, the rendering of the eyes changes. Faces look substantially different while retaining their original structure. Because of this, the argument has returned to a question that arose back in the spring, after the technology's first showing: critics say that the characters begin to look as if generated by a neural network.

Jensen Huang, the head of NVIDIA, has publicly explained that the result is still determined by the artist's work. Ways to adjust the faces do exist, but how far that corresponds to the intent of the game's authors is a separate question. An additional difficulty is that there are many screenshots and videos online and the authenticity of some of them cannot be verified.

DLSS 5 утекла до релиза: моддеры запустили нейрорендеринг NVIDIA в чужих играх и потеряли половину кадров — One of the comparisons that spread across the web: a character's face without DLSS 5 neural rendering and with it. Frame from a community video / via 80.lv

The assessment from developers

Daniel Vavra, creative director of Warhorse Studios, says he sees great potential in the technology's future.

Frames are lost even on suitable hardware

Enthusiasts' measurements show a noticeable price for the effect. In Control at 4K on a GeForce RTX 5070 Ti, enabling the mod drops performance from 71 to 35 frames per second, and on an RTX 5090 from 91 to 50. At lower resolutions the drop is expected to be substantially smaller. The leaked library supports only RTX 50 cards, but enthusiasts have tried running it on the RTX 30 generation as well — performance there is predictably low.

NVIDIA has not yet published an official list of supported graphics cards; a restriction to the RTX 50 line is considered the most likely. All the figures relate to an unfinished pre-release library launched by a mod in games without official integration — they cannot be carried over to the final version. What exactly the company is preparing for release and when the release will take place has not been specified publicly.

Why the argument about DLSS 5 is not only about the picture

NVIDIA announced DLSS 5 in March at GTC, calling it a real-time neural rendering model and promising a release in the autumn. In essence it is a third category after upscaling and frame generation: previous DLSS generations were sold as a way to raise the frame rate, the current one as a way to raise image quality at the cost of that very frame rate. The change of logic is notable, and it largely explains why the industry's reaction turned out to be so polarized.

Behind the leak lies a question that matters more to studios than any measurements: who is responsible for a character's final look. NVIDIA's earlier steps in this area were technically neutral — the upscaler restored what the engine had already drawn. The generative stage intervenes for the first time in what has traditionally been considered the area of responsibility of the character artist and the lighting artist. Some developers, judging by public reactions, see in it a way to get around the limitations of their own pipeline — above all in lighting and shadows. Another part of the industry is skeptical: if the picture can be touched up after the fact by a neural network, the requirements for how carefully the source assets have to be prepared change as well.

What happens next is fairly predictable. The leak has set an uncomfortable bar for NVIDIA: the release version will now also be compared with the early mods that have already spread across the web. Attempts to run the model on older cards add a second question — how strict the generational restriction will turn out to be and how the company will justify it. If exclusivity for RTX 50 is retained, interest in modified libraries on older graphics cards, judging by the community's activity, will probably not go anywhere.