# Orbis

A driving world model from the University of Freiburg that predicts long front-camera video rollouts from a few past frames, built to keep long-horizon predictions from degrading. Code and checkpoints are released under MIT.

Organization: University of Freiburg. Released: July 2025. Kind: Driving.

## What you control

- A few past frames from a car's front camera
- How many future frames to generate

## License

Weights: Available (https://huggingface.co/lmb-freiburg/Orbis). Declared license: MIT.
The weights' model card declares MIT, a permissive license that allows commercial use. Check what the model was built on, because a base model's terms can still apply.

## Real-time

No real-time claim.

## How to run it

1. Clone lmb-freiburg/orbis and create the conda environment from environment.yml.
2. Download the tokenizer and world model checkpoints from Hugging Face and place them in the experiment folders.
3. Run evaluate/rollout.py on the example input frames, or point a config file at your own frames.

## Where to try

No hosted destination verified.

## Sources

- [Project source](https://github.com/lmb-freiburg/Orbis)
- [Weights and model card](https://huggingface.co/lmb-freiburg/Orbis)
- [Paper](https://arxiv.org/abs/2507.13162)

## Questions

### Is Orbis open source?

Orbis publishes downloadable weights under MIT. The declared license is permissive.

### Can I use Orbis commercially?

The weights' model card declares MIT, a permissive license that allows commercial use. Check what the model was built on, because a base model's terms can still apply.

### Does Orbis run in real time?

The cited sources do not claim real-time operation.

### Where can I try Orbis?

No hosted version was verified; run the published weights yourself.

### What does Orbis take as input?

You give it a few past frames from a car's front camera and how many future frames to generate.

### How do I run Orbis?

Following the official instructions: Clone lmb-freiburg/orbis and create the conda environment from environment.yml. Download the tokenizer and world model checkpoints from Hugging Face and place them in the experiment folders. Run evaluate/rollout.py on the example input frames, or point a config file at your own frames.

### What are alternatives to Orbis?

Similar models in this directory include comma.ai world model 4B (open weights, permissive license), Oasis 3 (no public weights, service terms) and Orbis 2 (open weights, permissive license).

Checked 2026-11-03. Corrections: https://github.com/windflow-labs/world-model-directory