PAN 2 is an open-source AI world model from the Institute of Foundation Models (IFM) at MBZUAI. You give it one photo and describe what happens next, and it predicts the result as a video, with the camera move you ask for. The code and weights are released under Apache 2.0. In Promptus you can run it with one click from a CosyFlow, locally or in Playground. Watch our full PAN 2 test.
| Developer | IFM (Institute of Foundation Models), MBZUAI |
| Parameters | About 29 billion |
| Input | A start image plus scene, event and camera prompts (text to video is also supported) |
| Output | Video at 1248×704, 24 fps, clips up to 20 seconds per IFM |
| Licence | Apache 2.0 (code and weights) |
| ComfyUI | Native (core nodes); text encoder type set to pan2 |
| Model files | Diffusion model 26.7 GB (int8), Qwen3.5 9B text encoder 9.7 GB (int8), Wan 2.2 VAE 1.4 GB |
Most image-to-video models animate a picture. A world model is trained to predict how a scene changes over time. PAN 2 was trained on structured captions, so it works best when you describe the scene, then what happens next (one action per line, in order), then the camera move. Our workflow builds that caption from three prompt boxes.
On an RTX 5090 Laptop GPU (24 GB VRAM) with 32 GB of RAM, a 5-second clip at 1248×704 took 14 min 37 s. VRAM peaked at 23.7 GB and system RAM at 31.5 GB.
Prefer plain ComfyUI? Download the PAN 2 ComfyUI workflow or read our PAN 2 guide.
Sources: IFM's PAN 2 announcement and the PAN 2 model card. Speed and memory were measured by Promptus.
Join our distributed GPU compute network. Help us make AI accessible, scalable
and secure for designer, developers and start-ups.