An enthusiast has forced Nvidia’s DLSS 5 neural rendering into more than 50 games that were never designed to support it, ranging from a DOS version of Pong and the original Duke Nukem to The Witcher 3 and Marvel’s Spider-Man. Valery Kryzhanovsky documented more than 100 recordings from the experiment, testing how the technology behaves when deprived of the engine data and developer tuning Nvidia normally expects. His conclusion is that the underlying technology can produce striking images, but the forced implementation remains highly inconsistent in motion.
DLSS 5 officially debuted in NBA 2K27 in September as Nvidia’s 3D-Guided Neural Rendering technology. In a supported game, it receives information from the engine about objects, materials, lighting and motion, while developers can control where and how strongly the model changes the final image. Kryzhanovsky’s experiments deliberately remove much of that context, effectively stress-testing what the model does when it is applied to games it was never built to understand.
For older 2D games, Kryzhanovsky used DOSBox together with dgVoodoo, ReShade and RenoDX to place DLSS 5 into the output pipeline. The original 1991 Duke Nukem is one of the more unusual examples, with the neural model adding perceived depth, materials and lighting to the game’s simple 16-color sprites. Depending on the strength and number of processing passes, surfaces can begin to look embossed or three-dimensional, although stronger settings quickly produce distorted and unpredictable imagery.
Pong pushes that experiment even further because there is almost no visual information for the neural model to interpret. Rather than simply sharpening the minimalist scene, DLSS 5 attempts to infer more elaborate visual structure from what is essentially a collection of basic shapes. The result illustrates both the model’s ability to generate additional visual information and the danger of asking it to interpret imagery far outside the type of modern 3D scene it was designed around.
Kryzhanovsky found some of the most convincing results in games with largely static scenes. Heroes of Might and Magic, for example, produced comparatively consistent transformed imagery because the model had fewer rapid changes between frames to contend with. Once characters, cameras and environments begin moving, temporal inconsistency becomes much more obvious because this unofficial implementation lacks the motion information and developer controls used by proper DLSS 5 integrations.
The tests also exposed more serious problems in 3D games. Applying the model at different stages of the rendering pipeline could alter the apparent geometry of faces and objects, while details sometimes changed between consecutive frames. Kryzhanovsky found this especially noticeable while experimenting with RBDOOM-3-BFG, demonstrating why simply injecting neural rendering after the fact is fundamentally different from integrating it into a game engine.
That distinction is important when comparing the experiment with projects such as a GTA IV mod that adds native DLSS, FSR and HDR support. Traditional DLSS Super Resolution integrations rely on data such as motion vectors to reconstruct an image over time. DLSS 5 goes further by using generative neural rendering to alter lighting and material appearance, making the lack of accurate scene information considerably more consequential when the technology is forced into unsupported software.
Nvidia’s official implementation in NBA 2K27 is much more controlled. The company says DLSS 5 uses the rendered frame alongside engine data while allowing developers to select models, adjust structure and tone, and mask specific characters, objects or portions of a scene. Nvidia also says the technology is designed to preserve underlying geometry and maintain temporal consistency when integrated correctly, two areas where Kryzhanovsky’s unsupported experiments frequently break down.
The 50-plus-game test therefore should not be read as a benchmark of how properly implemented DLSS 5 will behave. Instead, it provides an unusual look at what the neural model attempts to reconstruct when stripped of the supporting information normally supplied by a modern engine. The results range from surprisingly coherent reinterpretations of old artwork to severe visual instability, showing both the capabilities of Nvidia’s new rendering approach and why developer-guided integration matters.
