NVIDIA announced advances in neural rendering, world models and AI simulation at SIGGRAPH 2026, including new Cosmos fou
NVIDIA presented its latest AI research and engineering developments at the SIGGRAPH conference in Los Angeles through July 23, 2026. The company's keynote on July 20 featured leaders Neil Ashton, Edward Liu and Ming-Yu Liu discussing neural rendering techniques, world models and simulation methods that are transforming how digital worlds are created across creative, industrial and robotics domains.
Edward Liu, director of applied deep learning research, showcased advances in 3D-guided neural rendering that address three key challenges: preserving artistic intent, ensuring temporal stability across video frames and rendering 4K content in real time. He emphasized that AI is extending graphics capabilities the way programmable shaders and ray tracing did previously, with simulation defining the world, generation enriching its appearance and artists directing the outcome.
Neil Ashton, distinguished engineer, covered innovation in AI physics with applications spanning weather forecasting, automotive aerodynamics and thermal design. NVIDIA's Earth-2 family of open models enables scientists to boost simulation resolutions using AI models trained on simulation data while achieving accuracy equal to or greater than traditional approaches. Ashton highlighted how latest model architectures compress model checkpoints to a couple hundred megabytes—a million-times compression—enabling physically accurate visualizations within a second.
Ming-Yu Liu, vice president of Cosmos Lab, announced the latest advancements in NVIDIA Cosmos, a world foundation model platform. The Cosmos 3 model uses a mixture-of-transformers architecture to provide global understanding across different embodiments—humanoid robots, grippers and self-driving cars—by building a common vocabulary for tasks including world understanding, prediction, simulation and action. NVIDIA released Cosmos 3 Edge today, a 4-billion-parameter model designed to run in real time on device. The company also announced Cosmos-Dreams, closed-loop simulators that can generate entire worlds from a single frame for autonomous vehicles, running on a single NVIDIA RTX PRO 6000 GPU. Developers can use these simulators to verify model accuracy before deployment or train on AI-generated scenarios difficult to create in the real world.
NVIDIA announced that leading creative applications are now opening Model Context Protocol connections that enable AI agents to work inside tools where scenes, shots, timelines and assets come to life while creators maintain control. Artists and technical directors can ask agents to inspect scenes for missing textures, flag color management inconsistencies, prepare export variants, generate playblasts or validate shots against pipeline rules. Adobe is expanding its creative agent across Firefly, Express and Creative Cloud, and providing the Adobe Express Developer MCP Server for AI coding assistants. Affinity by Canva introduced an AI Connector for Claude using MCP to bring natural-language automation into Affinity for repetitive tasks like renaming layers, resizing assets across channels and optimizing vector paths. NVIDIA RTX PRO workstations, DGX Spark and DGX Station systems are designed to run models and agents locally, improving responsiveness, reducing reliance on external services and keeping sensitive creative data in controlled environments.