Windflow

Memorizon

MBZUAI (Institute of Foundation Models) · Released October 2026

Memorizon trains a world model to remember beyond its context window, so long camera-driven videos stay consistent. You give it one image and steer with keyboard actions or a pose trajectory; it is built on Wan2.2-TI2V-5B and released under Apache-2.0.

Get the weightsCan I use it commercially?
Weights
Open weightsApache-2.0
Commercial use
Permissive license
Real-time
No real-time claim
Where to try
Run it yourself

01 Controls

What you control

  • A starting image
  • Keyboard actions
  • A camera pose trajectory

02 License

Can you use it commercially?

Permissive license

The weights' model card declares Apache-2.0, a permissive license that allows commercial use. Check what the model was built on, because a base model's terms can still apply.

Declared license: Apache-2.0 · Model card

A reading of the published license text, not legal advice.

03 Speed

Does it run in real time?

No real-time claim. 1280×704 at 16 fps, generated one second at a time by a 4-step distilled model.

Figures are the authors' own, with their settings and hardware. They are not comparable across models.

04 Setup

How to run it

  1. 01Clone TingtingLiao/memorizon, create a Python 3.11 conda environment and run pip install -e .
  2. 02Run python scripts/generate.py with the Luffuly/memorizon-1280x704-4step checkpoint and a starting image.
  3. 03Steer it with keyboard actions, one every 0.25 seconds, or replay a recorded trajectory.
  4. 04In Python, the checkpoint loads with MemorizonPipeline.from_pretrained.

Summarised from the project's own instructions. Follow the source for exact commands and versions.

05 Access

Where to try it

No hosted version was verified. Run the published weights yourself.

Listed alphabetically. Windflow maintains this directory and is one of the hosts; a listing does not promise capacity at any given moment.

06 Questions

Memorizon FAQ

Is Memorizon open source?

Memorizon publishes downloadable weights under Apache-2.0. The declared license is permissive.

Can I use Memorizon commercially?

The weights' model card declares Apache-2.0, a permissive license that allows commercial use. Check what the model was built on, because a base model's terms can still apply.

Does Memorizon run in real time?

1280×704 at 16 fps, generated one second at a time by a 4-step distilled model.

Where can I try Memorizon?

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

What does Memorizon take as input?

You give it a starting image, keyboard actions and a camera pose trajectory.

How do I run Memorizon?

Following the official instructions: Clone TingtingLiao/memorizon, create a Python 3.11 conda environment and run pip install -e . Run python scripts/generate.py with the Luffuly/memorizon-1280x704-4step checkpoint and a starting image. Steer it with keyboard actions, one every 0.25 seconds, or replay a recorded trajectory. In Python, the checkpoint loads with MemorizonPipeline.from_pretrained.

What are alternatives to Memorizon?

Similar models in this directory include ABot-World (open weights, permissive license), AlayaWorld (open weights, custom terms) and DreamX-World (open weights, permissive license).

Similar world models

All 56 in the directory