Nvidia reveals DLSS 5 details: three neural models and masks for artistic control

At GTC on March 16, Nvidia announced DLSS 5, and CEO Jensen Huang called it a 'GPT moment for graphics,' but the announcement drew heavy criticism, with users labeling the technology 'AI slop.' On July 20 at SIGGRAPH in Los Angeles, Nvidia finally detailed how it works. DLSS 5 includes three neural networks (model A, model B, model C) that differ in processing balance and resource demands. Developers can choose which models to use per scene and switch between them on the fly without latency, and the system uses temporal streaming to reduce flickering between frames. A key feature is masks and two intensity regulators (structural and tonal) that operate by object, allowing artists to disable generative processing on a character's face while keeping it on the environment. Classic rendering and upscaling still handle geometry, camera, composition, and basic light, with DLSS 5 as a generative layer on top. In March, the demo ran on two GeForce RTX 5090 cards, but Nvidia says optimized versions will fit on a single GPU; a 4K frame could take 16 milliseconds (60fps). Nvidia has not revealed minimum VRAM or supported GPUs, but the RTX 50 series (Blackwell) is considered the base. DLSS 5 is slated for fall 2026, with confirmed games including Starfield, Assassin's Creed Shadows, Hogwarts Legacy, and Oblivion Remastered, plus support from Bethesda, Capcom, and Ubisoft. Lead engineer Edward Liu admitted that true real-time photorealism has not yet been achieved.