CozyClay Debuts Open-Source 3D Previs Tool for AI Video

A new open-source previsualization tool called CozyClay aims to solve one of the most frustrating aspects of AI video generation: the lack of precise camera control. Instead of relying on unpredictable text prompts to describe complex camera movements, CozyClay provides a browser-based 3D staging environment. Creators can block scenes, position characters, and map out camera paths on a timeline before sending the structured spatial data directly to an AI video model.

Built on React Three Fiber and Three.js, the project is licensed under AGPL-3.0 and requires Node 22 or higher. Getting started is designed to be frictionless; running the command npx cozyclay launches a local editor at http://127.0.0.1:5180. Because the published npm package includes a prebuilt distribution folder, developers do not need to compile code or manage a complex dependency tree to begin staging their shots.

CozyClay integrates several advanced features to streamline the creative workflow. It ships with a Model Context Protocol server, allowing Anthropic's Claude to control the viewport using natural language commands. For character movement, the tool connects to NVIDIA ARDY for motion generation over an SSH-accessible GPU box. It also features a token-free ARDY text encoder configuration that bypasses the typical Hugging Face gate, simplifying the backend setup.

For 3D artists and filmmakers, the interface offers familiar Unity-style navigation. Users can fly through the scene with a right-drag, walk using WASD keys, orbit selections with Alt-drag, and frame objects by pressing F. Objects and primitives are manipulated using a standard W/E/R gizmo, while an overhead plan view helps map out character waypoints. By replacing vague text prompts with concrete 3D coordinates, CozyClay gives directors the exact framing they need for reliable AI video outputs.