Nvidia unveils AI tools for media, robotics & design
Tue, 21st Jul 2026 (Today)
Nvidia has announced a range of AI and graphics products and research updates at SIGGRAPH, spanning media verification, creative software, robotics and simulation.
The updates include a synthetic video detection service for newsrooms, a new edge world model for physical AI, local agent tooling for DGX Station systems, and broader support for Model Context Protocol connections in creative applications.
Media verification
One of the most directly commercial launches is Synthetic Video Detector, an Nvidia NIM microservice for editorial and media workflows. The tool analyses video frame by frame and produces a classifier score indicating whether footage may contain synthetic content.
Editorial teams can use that score to prioritise clips for review, flag questionable footage, or escalate material for deeper analysis. Nvidia presented the software as an additional signal within existing verification processes, not a replacement for newsroom checks.
According to Nvidia, the model remained effective after common processing steps such as compression, resizing, cropping and re-encoding. In Nvidia testing, accuracy reached up to 92% on uncompressed video, 87% at 15% compression, and 82% at 50% compression.
Processing speeds were given as little as 22 milliseconds for 1080p video on Nvidia RTX systems and about 30 milliseconds on Nvidia L40 GPUs. Organisations can deploy the software in on-premises, edge, hybrid and approved air-gapped environments where data control and residency are important, Nvidia said.
Wowza is embedding the microservice through its Video Intelligence Framework. Nvidia said this brings synthetic video detection into livestreaming operations already used across more than 35,000 deployments in over 170 countries.
Creative software
Another set of announcements centres on Model Context Protocol, or MCP, which allows AI agents to work inside creative and production software. Nvidia said the approach lets artists and technical teams use agents for tasks such as checking scenes for missing textures, validating shots against pipeline rules and preparing exports, while leaving creative decisions to users.
Nvidia Agent Toolkit supports MCP integration through both an MCP client for connecting to remote servers and an MCP server for publishing tools to MCP clients.
Several software groups are exposing MCP connections or MCP-ready workflows. Adobe is expanding its creative agent across Firefly, Express and Creative Cloud, while also offering an Adobe Express Developer MCP Server for coding assistants building Express add-ons.
Affinity by Canva has introduced an AI Connector for Claude using MCP. Nvidia said the setup allows designers to automate repetitive tasks such as renaming layers and artboards, resizing assets for different channels, applying bulk edits and preparing files for delivery.
Blender, Boris FX Silhouette, Foundry Griptape, SideFX Houdini and Unreal Engine were also cited as examples of software adopting MCP links. The common theme is closer connections between AI agents and production context in applications used for design, visual effects and game development.
Edge models
Nvidia also said Cosmos 3 Edge is now openly available. The 4-billion-parameter model is designed for deployment on Jetson, RTX Pro, DGX and GeForce RTX hardware, and is intended for physical AI uses such as robotics, smart infrastructure and autonomous systems.
According to Nvidia, the model can understand and generate text, image, video, ambient sound and action. Nvidia said its architecture supports on-device vision analytics and robot action in real time.
Partner companies including Agile Robots, Doosan Robotics, Siemens and Skild AI are evaluating the model for robotics work, while Centific, Vaidio and YUAN are assessing it for vision agents at the edge. Nvidia said developers can post-train the model on proprietary robot and sensor data using DGX Station and then deploy it on Jetson Thor systems.
The launch adds to the broader Cosmos range, which Nvidia said is available in Edge, Nano and Super sizes for different stages of development and deployment.
Desktop agents
Nvidia also outlined a local AI agent stack for DGX Station. Nvidia Agent Toolkit on DGX Station combines NemoClaw, the Nemotron 3 Ultra open model, Omniverse libraries and a secure runtime on a local machine without requiring internet access.
Users can link multiple systems together as workloads grow, Nvidia said. It also named LangChain, Nous Research and OpenClaw among groups tuning or using the stack for agent workflows.
The hardware is available to order from ASUS, Dell Technologies, Exxact, GIGABYTE, HP, MSI and Supermicro.
Research focus
Alongside the product launches, Nvidia highlighted 21 accepted technical papers at SIGGRAPH. The company said the work focuses on real-time systems for generating virtual worlds and supporting machine training in physical settings.
Examples include MotionBricks, a real-time motion model trained on more than 350,000 motion clips, and GPC, a framework for training generative controllers on large-scale motion data. Nvidia also pointed to ArtiFixer for cleaning up 3D captures into virtual scenes, a new solver for materials such as snow and sand in the Newton physics engine, VideoNeuMat for extracting reusable materials from generative video models, and ARDY for steering 3D character motion from text prompts.
The breadth of the announcements shows Nvidia pushing the same underlying strategy across multiple markets: bringing AI models and agents closer to production tools, edge devices and local infrastructure, while extending its reach from graphics into media operations, robotics and industrial systems.
The papers are openly available, Nvidia said, with code and models free to download.