WorldCrafter: Video Generator — What It Does and What You Need to Run It

WorldCrafter is an open source video generator from ARC Lab, Tencent PCG. It tackles image-to-video: turning a still image into a video sequence. The developers describe the project as a "Consistent Video World Model with Implicit 3D-aware Memory" — a world model for video with frame consistency and implicit memory that accounts for the three-dimensionality of the scene.

What it does

The main task is generating video from an image. The authors position WorldCrafter as a world model: it doesn't just animate a picture, it aims to maintain scene consistency across frames. To do this, it uses implicit memory with 3D information taken into account. The authors do not specify architectural details or concrete limits on video length or resolution in the fact sheet.

The code is written in Python and built on the diffusers library. The repository is tagged with video-generation and world-models.

What you need to run it

It requires an NVIDIA GPU with a compatible driver. As described by the author, the platform is CUDA, and the environment is pinned to PyTorch 2.10 and CUDA 12.8. The Python version is 3.11, and it is intended to run on Linux. Installation can be done via uv or conda with pip.

The largest model weights file is 9.3 GB, and the full set of all model weights files is 137.1 GB. This is worth keeping in mind when planning disk space.

The code license is the License Term of WorldCrafter. The model may be used, copied, modified and distributed for academic purposes only. Any commercial or production use is prohibited. When distributing, you must include the copyright notice and the text of the permission. Third-party components retain their original licenses, and their terms must also be observed.

The repository is available at github.com/TencentARC/WorldCrafter, the model at huggingface.co/TencentARC/WorldCrafter-Fast, and the project site at drexubery.github.io/WorldCrafter.

Who it suits

WorldCrafter is aimed at research tasks in video generation and world models. It will be useful to those studying video sequence consistency, implicit memory and 3D context in generative models. The availability of open source code and published weights makes it possible to understand the implementation and adapt it for your own experiments.

The tool is not suitable for commercial projects or production use: the license explicitly prohibits this. You should also take hardware requirements into account — an NVIDIA GPU with CUDA 12.8 support is needed, and the full weight volume exceeds one hundred gigabytes, which may be excessive for quick experiments.

WorldCrafter is a research tool with a clear niche: academic work on video world models. It requires a serious GPU and significant disk space, and its license restricts use to academic bounds.