WorldCrafter is a video world model from ARC Lab at Tencent IEG with Peking University, described in the paper "WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory" (arXiv 2609.24984). It enables camera-controlled scene exploration from a single image or a text prompt, and its camera-queryable implicit 3D-aware memory is meant to preserve scene information across viewpoints and over long horizons. Two checkpoints are published on Hugging Face: WorldCrafter-Base, and WorldCrafter-Fast, a distilled pair of high- and low-noise models for faster inference. The Base model reuses shared components from the Fast download, so both folders are needed to run it. The inference.py script runs image-to-video and text-to-video from a camera path file, with an optional automatic prompt writer for custom images that uses Qwen3-VL-4B-Instruct in first-person or third-person style. Torch compilation is off by default and enabled with --enable-compile. The repository also ships interactive demo code and serving infrastructure: running python -m demo with the Fast weights opens a local web page on port 8080 with keyboard camera controls, and --devices 0,1 spreads it over two GPUs. A hosted demo Space is linked from the README. Setup targets Linux with Python 3.11 and an NVIDIA GPU with a compatible driver, using a locked PyTorch 2.10 and CUDA 12.8 environment through uv or conda plus pip, with FFmpeg and FlashAttention 3; the README does not state a GPU memory requirement. The license is a custom Tencent term that grants broad rights but conditions them on using WorldCrafter only for academic purposes and never for commercial or production purposes; third-party components such as Helios-Base remain under their own licenses.