NVIDIA’s DLSS technology is about to go pretty nuts. Since the beginning, DLSS has been mainly focused on reconstructing images, improving ray-traced visuals, or generating frames. However, DLSS 5 is designed to use generative AI to alter rendered imagery. The technology isn’t officially available yet, but an early build has already been leaked into the hands of modders.
New look at DLSS 5 with the leaked files in games like Silent Hill, Wuchang and The Last of Us pic.twitter.com/q9seI1ILLs— Pirat_Nation 🔴 (@Pirat_Nation) August 28, 2026
The leak came from an early-access build of NBA 2K27, which contained an NVIDIA DLL that included DLSS Neural Rendering, or DLSS-NR. The file appears to be an unfinished version of the technology NVIDIA plans to release later this year. Despite not being officially supported, modders have already managed to make it run in games including Control, GTA 5, and Skyrim.
DLSS 5 Is Already Changing Faces in GTA 5 and Control
Image Credit: NVIDIA
Image Credit: NVIDIA
Image Credit: NVIDIA
Image Credit: NVIDIA
Image Credit: NVIDIA
The leaked nvngx_dlssnr.dll is considerably larger than previous DLSS libraries, weighing in at around 158 MB. Analysis of the file suggests that most of its size comes from the neural network itself, with approximately 140 MB dedicated to model weights. The discovered build also appears to be designed for RTX 50-series GPUs, although that doesn’t necessarily mean the final version will remain exclusive to Blackwell hardware.
Apparently modders have managed to enable DLSS 5 AI Neural Filters in Control, using a Neural Rendering DLL from NBA 2K27 early access. It seems to be limited to RTX 50 Series for now.Performance impact is about 40% (RTX 5060 drops from 86 fps to 34 fps with NR enabled) pic.twitter.com/3UQ4XfGmLt— Alex (@PlanoGaming) August 28, 2026
The RenoDX modding community has been experimenting with the library, and Control has become one of the clearest demonstrations. In its current form, the neural rendering effect is primarily being applied to character models. Users can switch between several visual styles and adjust the intensity of the transformation.
That can produce dramatically different results. Jesse Faden’s face can become smoother, more detailed, and more conventionally photorealistic, but pushing the effect too far can also make her look noticeably different from the original character. Similar experiments have appeared in GTA 5, including footage showing Michael receiving the unofficial AI treatment.
The effect has also been tested in Skyrim, where the improvement to character faces can be particularly obvious because of the game’s aging character models. The modded implementation can make faces appear substantially more realistic while also changing the game’s established visual style.
However, this isn’t a free upgrade. In one Control test, an RTX 5070 Ti reportedly saw performance fall from around 71 frames per second to roughly 35 FPS with the experimental implementation enabled.
The Leaked Build Is Not Representative of Final DLSS 5
There is an important caveat surrounding virtually every DLSS 5 video that you can currently find online: this isn’t the finished product.
The leaked DLL is signed by NVIDIA and was built recently, but analysis indicates that it is still incomplete. It also wasn’t designed to simply be dropped into arbitrary games. Developers need to integrate the technology and provide the required inputs for the neural rendering system.
Interestingly, the DLL appears to require fewer inputs than NVIDIA’s Ray Reconstruction technology. It can work with elements such as the rendered image, depth, motion vectors, and UI information, suggesting that DLSS 5 operates late in the rendering pipeline. That could help explain why modders have been able to adapt it to existing games relatively quickly.
For now, though, the biggest question is whether NVIDIA can turn this experimental technology into something that actually benefits us gamers. What do you think will happen next? Let us know in the comments!
This post belongs to FandomWire and first appeared on FandomWire
