nano-world-model: video generator — what it does and what you need to run it

nano-world-model is an open-source video generator that the authors position as a minimalist toolkit for research on world models. The repository is built for tasks where you need not just to generate video, but to work with a model that predicts how a scene will evolve, including for planning and robot control.

What it does

As described by the authors, the project is a "minimalist repository with batteries included for advancing world model science." The repository tags include:

• video generation and streaming video;
• diffusion models and diffusion-forcing;
• model-predictive control;
• planning;
• robot manipulation.

The authors do not provide a detailed list of supported generation scenarios or limits on video duration and resolution. The model referenced by the repository is tied to the point-maze task, which points to work with trajectories and spatial scenes rather than arbitrary text descriptions of a plot.

What you need to run it

The code is written in Python using the PyTorch library. The code license is MIT. According to the author's description, a CUDA platform is required; the documentation mentions "CUDA wheels and all dependencies," with a reference to the environment setup section.

The weight files take up 0.6 GB — this is the total size of all model files, and the largest single file also weighs 0.6 GB. The model is hosted on Hugging Face. The authors do not specify minimum VRAM requirements or specific CUDA and PyTorch versions.

Who it's for

The project is aimed primarily at researchers working on world models, diffusion models, and reinforcement learning. It will be useful to those who need compact open-source code for experiments with video generation in the context of planning and robotics, rather than a ready-made user-facing tool.

For those looking for a text-to-video generator with a polished interface and predictable quality on arbitrary plots, this repository is unlikely to fit: the authors do not describe it as an application service, and the emphasis is on research minimalism and working with a specific class of tasks.

nano-world-model is an open-source research project that provides a starting point for experiments with world models and video generation in planning tasks. Its value is defined not by the breadth of user scenarios, but by the availability of the code and the ability to study and modify the approaches the authors have laid out.