Nvidia DLSS 5 at SIGGRAPH | Summary

What if a game could keep everything its creators designed but make every frame look more like reality? At Siggraph in July 2026, Nvidia introduced DLSS 5, a new approach that combines traditional rendering with real-time generative AI. Its central idea is simple: simulation defines the world, generation enriches its appearance, and artists direct the result. Traditional graphics construct an image from authored geometry, materials, lighting, and physics. Every object and camera angle is deliberately controlled. Generative models offer something different. After learning from images of the real world, they can reproduce subtle details such as light scattering through skin, reflections in glass, or sunlight passing through hair. But those models are usually large, slow, and unpredictable. DLSS 5 attempts to combine the strengths of both systems. The game engine renders the world as designed. Then a specialized generative model enhances that rendered frame's final appearance in real time. The result can include richer materials, more convincing reflections, improved contact shadows, better environmental lighting, and more realistic skin, hair, and foliage. But making an image look better is only half the challenge. It must also remain faithful to the original game. A general image model might alter a character's identity, remove an intentional scar, or reinterpret part of a scene. That may be acceptable during offline image generation where several results can be reviewed. A game doesn't have that luxury. Every generated frame is shown immediately. To constrain the model, DLSS 5 uses more than the visible image. It can also receive information from the renderer, including surface colors, normals, lighting data, and other internal buffers. These signals tell the model what must remain consistent while still allowing it to enhance the qualities that make the image feel real. The second challenge is time. Most video generators process several frames together. A game must respond instantly to player input, so DLSS 5 operates causally. One frame enters and one enhanced frame comes out. Motion vectors from the game engine help it understand how objects move between frames, reducing shimmer, drifting, and visual instability. The third challenge is speed. Instead of running a massive general-purpose model, Nvidia distills its visual knowledge into a smaller, highly specialized one-step model. It can't generate anything imaginable. It has one focused job: make real-time rendering look more realistic while preserving the experience of playing a game. Developers and artists retain control over that process. They can choose different enhancement models for different scenes, adjust structural detail and overall tone, isolate characters automatically, or use engine-provided masks to control individual objects. A character, a bottle, a group of props, or an entire environment can each receive a different level of enhancement. The goal is not to let the AI make the artistic decisions. It's to give artists another tool for expressing those decisions. Nvidia describes this as a third category of neural graphics. Earlier DLSS systems primarily reconstructed missing detail. Other neural techniques approximate expensive rendering calculations. DLSS 5 adds generation where the model actively contributes to the final appearance. It doesn't replace rasterization, ray tracing, or path tracing. It builds on top of them. Better source rendering still gives the model better information to work from. Nvidia also acknowledges that real-time photorealism hasn't been solved. Appearance currently benefits more than animation and greater generative capability will require even stronger creative controls. DLSS 5 represents Nvidia's vision for the next stage of computer graphics, authored worlds enhanced by generative intelligence without surrendering artistic intent. Not AI replacing graphics, AI extending what graphics can become.