Tencent ARC Releases WorldCrafter: Open-Source Video World Model with Implicit 3D Memory

Tencent ARC Lab has released WorldCrafter, a consistent video world model designed for camera-controlled scene exploration from images or text prompts. The model utilizes an implicit 3D-aware memory to preserve scene information across viewpoints and long horizons. Two variants are available: WorldCrafter-Base and WorldCrafter-Fast, the latter being a distilled model optimized for faster inference. The project's interactive demo code and serving infrastructure are fully open source, allowing users to run and customize the model locally using Python 3.11 and NVIDIA GPUs. The setup supports both uv and conda environments, and the interactive demo runs on localhost:8080 with keyboard camera controls.