PAN 2 World Model

PAN 2 World Model

Turns a text prompt, or a prompt plus a first frame, into a short simulated video with the PAN 2 world model. Aimed at people who want to try or study world model video generation.

Made by akhaliq on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space lets you try PAN 2 through a Gradio page. PAN 2 is a world model with 29B parameters, built at MBZUAI's Institute of Foundation Models. You describe what is happening, optionally with a first frame, and it simulates the outcome as a video of up to 20 seconds at 704 by 1248 and 24 fps. The demo runs the released perceptual backbone, a latent diffusion model, through the Diffusers pipeline.

Who it suits: It suits people who want to see how a world model predicts what follows from a described action, and developers who want to run the same pipeline in their own code.

Industry | Who uses it | How they use it | Example
Research | AI researchers | Study how a world model simulates the result of a described action | Describe an action and look at the predicted outcome clip
Video and media | Video creators | Try short generated clips from text or from a first frame | Turn a still frame plus a caption into a short clip
Software | Developers | Run the Diffusers pipeline from their own scripts | Load the model and export the result as an mp4 file

Skip it if you want long, full-quality video without a huge GPU, since the shared hardware only fits short clips at reduced resolution.

Where to use it:

• A browser: open the Space's page on Hugging Face, with nothing to install
• A copy of the Gradio app on Windows, macOS or Linux, using Python 3.12
• A Python script that uses the Diffusers pipeline on a machine with a large CUDA GPU
• A dedicated a100-large Space on Hugging Face, which the Space's notes suggest for unrestricted use

Use case | Who | What they get back
Text to video | Anyone with a scene in mind | A short clip simulated from the written description
Image to video | People with a starting picture | A clip that continues from the first frame you give
Action outcomes | Researchers | A predicted result for an action you describe in text

How to set it up: The online copy needs nothing. A copy on your own computer needs a lot of GPU memory, since the model takes about 85 GB in bf16.

• Python 3.12 with pip on Windows, macOS or Linux
• The Hugging Face hub client, which the first step installs
• The packages listed in requirements.txt

1. Install the Hugging Face hub client.
2. Download the Space's files into a local folder.
3. Move into that folder.
4. Install the Gradio version the Space pins.
5. Install the remaining packages from requirements.txt.
6. Start app.py.

pip install -U huggingface_hub
hf download akhaliq/PAN2 --repo-type space --local-dir PAN2
cd PAN2
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: In the Space, type a prompt, add a first frame if you want to start from a picture, and run the generation. In your own Python code, load the model with the Diffusers pipeline as below.

import torch
from diffusers import ModularPipeline
from diffusers.utils import export_to_video

pipe = ModularPipeline.from_pretrained("IFM/PAN2")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

video = pipe(
    prompt="A red fox trots through fresh snow in a pine forest at dawn.",
    height=704, width=1248, num_frames=121, num_inference_steps=50,
    generator=torch.Generator().manual_seed(123),
    output="videos",
)[0]
export_to_video(video, "pan2.mp4", fps=24)

Tool | What it does
Text-to-video | Makes a clip from a written prompt alone
Image-to-video | Makes a clip from a prompt plus a first frame

Example: You type "A paper boat drifts down a rain gutter while the camera follows from behind". The Space returns a short video that simulates that scene as a prediction of what would happen.

For newcomers and experts: If you are new, open the Space and write two or three lines about a scene: who does what, when, and where the camera moves. Start with a short clip, since long ones go over the shared GPU budget.

If you know the field, prompts work best as structured, time-grounded captions that bind timed actions to actors and note camera motion. The full PAN 2 workflow adds a separate conceptual reasoner that rewrites free text into that form, and this demo runs only the perceptual backbone. The Space asks ZeroGPU for a full RTX Pro 6000 and allows up to 300 seconds per generation.

You give a prompt and an optional first frame
The guardrail screens the prompt
The latent diffusion backbone simulates the outcome
The guardrail screens the generated video
You get a clip of up to 20 seconds

Good to know: The clips are predictions, not records of real events, so say they are AI made where they could mislead.

• License: apache-2, listed for both the Space and the model
• Cost to run: no API key is named, but unrestricted use needs a dedicated a100-large Space, and ZeroGPU use is capped by a daily quota
• Data it reaches: your prompt and optional first frame, and the generated video, all of which the PAN 2 guardrail screens
• Not for: the uses the model card lists as out of scope, and long full-quality generations on the shared hardware

On Hugging Face, 13 November 2026

• Likes: 3
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/akhaliq/PAN2
• Files: https://huggingface.co/spaces/akhaliq/PAN2/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.