PAN 2 AI World Model: Free ComfyUI Workflow Download

Published on:
November 17, 2026
Download the PAN 2 ComfyUI workflow (JSON). Run the open-source AI world model from IFM at MBZUAI locally: models, prompts, settings and measured VRAM.
PAN 2 AI World Model: Free ComfyUI Workflow Download

Quick answer: PAN 2 is a new open-source AI world model from IFM at MBZUAI (29B parameters, Apache 2.0). You give it a start image plus a short scene, event and camera description, and it predicts what happens next as a 1248×704 video at 24 fps. It runs locally in ComfyUI on core nodes. Download our ready-made image-to-video workflow below.

Download the PAN 2 ComfyUI workflow (JSON)

👉 Download promptus-pan2-image-to-video-comfyui-workflow.json

If the file opens in your browser instead of downloading, right-click the link and choose Save link as…. Then drag the .json file onto the ComfyUI canvas (or the ComfyUI tab inside Promptus) to load it.

Prefer one click? The same workflow is published as a Promptus CosyFlow: PAN 2 CosyFlow. From its ⋮ menu you can choose Install Locally (it downloads the models for you) or run it in Playground.

What is PAN 2?

PAN 2 is a "world model": instead of only animating a picture, it is trained to predict how a scene changes over time, so the action you describe plays out with the camera move you ask for. It was released by IFM (Institute of Foundation Models) at MBZUAI under the Apache 2.0 licence. Read the official announcement on the IFM blog.

Models you need

All three files come from Comfy-Org/PAN2 on Hugging Face. Put each one in the matching ComfyUI models folder:

  • pan2_int8_convrot.safetensors (about 26.7 GB) → models/diffusion_models
  • qwen3.5_9b_int8_convrot.safetensors (about 9.7 GB) → models/text_encoders
  • wan2.2_vae.safetensors (about 1.4 GB) → models/vae

You need a recent ComfyUI: update it first (we tested on ComfyUI 0.45.0). In the text encoder loader the type must be set to pan2; the workflow already does this.

How to use the workflow

  1. Start image: load the first frame of your scene.
  2. Scene prompt: describe what is in the picture and where (the subject, setting and lighting).
  3. Event prompt: describe what happens next, one action per line, in order.
  4. Camera prompt: describe the camera move (for example a slow push-in or a tracking shot).
  5. Press Run. The workflow joins the three prompts into the structured JSON caption PAN 2 expects, samples, and saves an MP4.

PAN 2 is trained on structured JSON captions, so a plain one-line prompt gives weaker results. That is why the workflow splits the prompt into these three boxes and builds the JSON for you. A caption guide note is included inside the workflow.

Settings used

  • Steps: 50
  • CFG: 2.5
  • Shift: 9
  • Output: 1248×704, 24 fps

Hardware: what we measured

On an RTX 5090 Laptop GPU, one clip took 14 min 37 s, with peak VRAM of 23.7 GB and peak system RAM of 31.5 GB. Plan for a 24 GB-class GPU and 32 GB+ of RAM, and about 38 GB of disk for the models. If your PC can't run it, the CosyFlow can also be run online in Promptus Playground.

Watch the full tutorial

The video walks through installing PAN 2 in ComfyUI, where each prompt goes, real example outputs, and running the CosyFlow in Promptus.

Video chapters

  • 0:00 PAN 2: new open source AI world model
  • 0:41 What is an AI world model?
  • 0:59 AI world model examples (PAN 2)
  • 1:31 PAN 2 image to video examples
  • 1:54 PAN 2 system requirements
  • 2:14 How to install PAN 2 in ComfyUI
  • 3:01 PAN 2 blurry output fix
  • 3:22 PAN 2 ComfyUI prompt guide (scene, event, camera)
  • 4:23 PAN 2 VRAM and speed test (RTX 5090 laptop)
  • 5:06 Run PAN 2 locally with a Promptus CosyFlow
  • 5:49 PAN 2 settings and limitations
  • 6:24 Is PAN 2 worth it?
View more workflows
promptus ai
Start running your first workflow
Go from idea to production-ready output in minutes.
Try Promptus for free ➜