NVIDIA has once again showcased its DLSS 5 neural rendering at the SIGGRAPH 2026 event, currently taking place in Los Angeles. Despite the initial negative reaction from gamers, NVIDIA is continuing to develop the technology, offering improvements and a new demonstration of how DLSS 5 is intended to function. One of the most interesting aspects is that game developers can choose from three DLSS 5 models, each providing a different level of structural intensity, global illumination, texture detail, and more. These models, labeled Model A, Model B, and Model C, can be tailored to specific scenes or character needs, allowing developers to mix and match according to their requirements. Each model carries a different parameter count, meaning that some are more compute and memory-intensive than others. However, they generally fit and run fine on a single GPU without any problems.
Another noteworthy feature is the preservation of artistic intent, which was heavily criticized in the original DLSS 5 demo. Each individual object, scene, and game setting can be customized according to the developer's vision. This is one of the challenges NVIDIA faced with DLSS 5, as maintaining visual identity, object semantics, and character poses, along with output lighting semantics, motion, and art direction, is crucial. Within these strict parameters, DLSS 5 applies material photorealism to enhance visual appeal. Below you can even see the customization of the applied model, where there are adjustable parameters that can be done per scene and per character.