NVIDIA Unveils DLSS 5 and Cosmos AI Push at SIGGRAPH Keynote

NVIDIA Unveils DLSS 5 and Cosmos AI Push at SIGGRAPH Keynote NVIDIA unveiled DLSS 5 , a new AI rendering technology aimed at making real-time graphics more photorealistic while preserving artistic intent and keeping game performance within tight limits. The company said it is shipping this fall . The keynote highlighted NVIDIA’s broader push to blend graphics, simulation and AI, including AI-assisted physics modeling that can make complex simulations dramatically faster and cheaper. NVIDIA cited use cases in weather, climate and aircraft engineering. NVIDIA also expanded its Cosmos platform for physical AI, announcing Cosmos 3 Edge for real-time on-device robot and video analytics, plus Cosmos Dreams for closed-loop simulation and training of autonomous systems. NVIDIA used its 2026 SIGGRAPH Research Keynote to outline a broad push to combine computer graphics, simulation and artificial intelligence, including a new DLSS 5 technology for real-time rendering, advances in AI-assisted physics simulation and new additions to its Cosmos world foundation model platform for physical AI. The keynote opened with NVIDIA framing computer graphics as entering “a new era,” with AI increasingly tied to rendering, simulation, robotics and digital twins. Jensen, who introduced the session, said NVIDIA’s history at SIGGRAPH has included programmable GPUs, CUDA, RTX and Omniverse, and argued that virtual worlds will be central to training robots before they operate in the real world. “Before robots operate in the real world, they will learn in virtual worlds with synthetic experiences,” Jensen said. “That is why computer graphics matter more than ever.” DLSS 5 Targets Real-Time Photorealism Edward Liu, NVIDIA’s Director of Applied Deep Learning Research and the technical leader behind DLSS, introduced DLSS 5, describing it as a new generation of the company’s AI rendering technology. Liu said DLSS 5 uses traditional rendering as a foundation, then applies generation to enrich the final appearance of the image in real time. “The renderer keeps building the world exactly as the game has authored it,” Liu said. “The generation becomes the learned stage afterwards to enrich its appearance.” Liu said DLSS 5 is intended to combine the controllability of rendering with the photorealistic knowledge learned by generative models. He emphasized that the technology is not designed to replace graphics pipelines, but to extend them. He described DLSS 5 as adding a third category of AI use in real-time rendering, alongside reconstruction and function approximation. According to Liu, NVIDIA had to address three core challenges: preserving artistic intent, maintaining temporal coherence frame by frame and fitting within the tight performance budget of real-time games. He said the model uses renderer outputs and internal buffers such as albedo, surface normals and lighting information to preserve details that are important to a scene, while enhancing elements such as subsurface scattering, material response, contact shadows and environment lighting. Liu said DLSS 5 runs causally, “one frame in, one frame out,” without looking ahead, and was distilled into a smaller one-step pixel-space diffusion transformer model focused specifically on making real-time rendering appear more realistic. He said DLSS 5 is “shipping this fall.” Artists Get Controls Over AI-Enhanced Frames Gaff, described as a creative artist, demonstrated how developers and artists can direct DLSS 5. He said the technology respects the original rendered frame and does not change geometry, but can uplift images by improving cont...