PAN 2 in ComfyUI: Free 20-Second AI Video, Apache 2.0

PAN 2 in ComfyUI: Free 20-Second AI Video, Apache 2.0

In qualità di Affiliato Amazon io ricevo un guadagno dagli acquisti idonei.As an Amazon Associate I earn from qualifying purchases.En calidad de Afiliado de Amazon, obtengo ingresos por las compras adscritas que cumplen los requisitos aplicables.

PAN 2 in ComfyUI is the kind of news that makes your head spin: a free video model, Apache 2.0 licensed, that generates clips up to 20 seconds long from text or an image. It was built by the Institute of Foundation Models at MBZUAI, the Abu Dhabi university devoted to artificial intelligence, and since November 15 ComfyUI supports it directly, with no external nodes. Version v0.46.0, released on November 18, lists it among the official changes.

Before you rush to download it, here is how we see it. We work locally on a perfectly ordinary RTX 3060 with 12 GB and 32 GB of RAM, and PAN 2 is a 29-billion-parameter model: the lightest ComfyUI file weighs 26.7 GB. We haven't tried it, and going by the numbers, getting it to run on a card like ours is very hard. It's still a model worth knowing, because it works differently from other video generators and its license is one of the most open around. Below you'll find what's confirmed, what you actually need and how to write the prompt.

What it isA "world model" that generates video from text or a starting frame, up to 20 seconds at 704×1248 and 24 fps
Who made itThe Institute of Foundation Models at MBZUAI (United Arab Emirates)
VRAM requirednot stated; the int8 ComfyUI file weighs 26.7 GB, plus 9.75 GB for the text encoder
Does it run on an RTX 3060 12 GB?not stated; on paper it's very hard, and there is no lighter GGUF version yet
LicenseApache 2.0 for code and weights; the Wan 2.2 VAE and Qwen3.5-9B have their own licenses
Commercial use from Italyyes, the Apache 2.0 license sets no commercial or geographic limits
Where to download itHugging Face: the official IFM/PAN2 repository and the Comfy-Org/PAN2 repack
Free or paidfree, it runs on your PC

What is PAN 2 and why is it called a "world model"?

PAN 2 is a video generator that "thinks" about the scene first and then draws it. You give it an action, for example "the cat jumps on the table and knocks over the glass", and optionally a starting frame. A language model (Qwen3.5-9B, frozen and not retrained) rewrites the request as a structured description of the state of the world: what happens, in what order, who performs each action, how the camera moves. Then a 29-billion-parameter MMDiT network (an architecture that processes text and image together) generates the video, and the Wan 2.2 VAE decodes it (the VAE is the part that turns the model's compressed data into actual frames).

That's where the term "world model" comes from: the goal is to simulate what happens after an action, with believable physics, more than to produce a pretty shot. According to the official Hugging Face model card, the standard output is 704×1248 at 24 frames per second: 241 frames (about 10 seconds) by default, up to 481 to reach 20 seconds. The authors are honest about the limits: roughly two out of three action plans contain at least one step the model doesn't carry out, and in long sequences the look of the subjects tends to drift.

How much VRAM does PAN 2 need in ComfyUI?

There is no official VRAM requirement (VRAM is the memory on your graphics card), but the files speak for themselves. The Comfy-Org repack includes these versions:

  • pan2_int8_convrot: 26.7 GB, the int8 compressed version (8-bit integers, lighter than the original);
  • pan2_bf16: 67.4 GB, full quality;
  • qwen3.5_9b_int8_convrot: 9.75 GB, the text encoder (19.3 GB in bf16);
  • wan2.2_vae: 1.4 GB.

The work done for the core integration cuts compute memory quite a bit: with 241 frames the activation peak drops from 12.5 to 3.8 GB, according to ComfyUI Wiki's article on native support. The 26.7 GB model is still there, though, and on a 12 GB card it has to be loaded in pieces from system RAM, which also has to hold the text encoder. With 32 GB of RAM we're at the limit, so we don't recommend it on setups like ours. There's also a community repack in NVFP4 format at 16.3 GB, but NVFP4 is a format designed for NVIDIA's RTX 50 series cards: if you're weighing an upgrade, our guide to graphics cards for AI compares the models and price ranges.

On speed, the only published figure is for an RTX 5090: 15.2 seconds per step with int8 weights and attention, 26.8 seconds per step dense, at 241 frames. With the recommended 50 steps that's about 12-13 minutes of sampling alone for a 10-second clip, on the most powerful consumer card there is. On smaller cards the times grow a lot.

How do you use PAN 2 in ComfyUI without external nodes?

You need a ComfyUI build newer than v0.45.0, then you build the graph with standard nodes, because there's no official template yet. The chain listed by ComfyUI Wiki is this:

  1. Load Diffusion Model (UNETLoader) with pan2_int8_convrot.safetensors in the models/diffusion_models folder;
  2. Load CLIP with qwen3.5_9b_int8_convrot.safetensors and the type set to pan2;
  3. CLIP Text Encode for the prompt;
  4. to start from an image, the Wan22ImageToVideoLatent node, the same one used by Wan 2.2, which uses the photo as the first frame;
  5. KSampler with 50 steps, euler sampler, simple scheduler, CFG 2.5 and shift 9 (the values from the original code);
  6. VAE Decode with wan2.2_vae.safetensors and Create Video at 24 fps.

If this is your first node graph, start with our beginner's guide to ComfyUI. After updating, if you get red nodes or out-of-memory errors, the ComfyUI errors guide explains how to fix them. Alternatively there's the community pack ComfyUI-PAN2, released on November 11, which adds a PAN2 Prompt Builder node to fill in the prompt field by field.

Why doesn't a one-line prompt work with PAN 2?

Because the model was trained on structured descriptions, not short sentences. ComfyUI Wiki reports that a one-line prompt, even the one in the official example, returns washed-out clips or noise. You need to give six pieces of information: the scene, the event, the subject, the lighting, the viewpoint and the style. The exact field names aren't documented, so the safest way is to use the Prompt Builder from the ComfyUI-PAN2 pack, or to write an orderly text like this example of ours:

Scene: a small wooden kitchen in the morning, a glass of water on the table.
Event: 0-4 s the cat walks to the table; 4-7 s it jumps on the table; 7-10 s the glass tips over and water spills.
Subject: an orange tabby cat.
Lighting: warm sunlight from the window on the left.
Viewpoint: static camera at table height, medium shot.
Style: realistic, natural colors.

Actions with time ranges help the model follow the order of events. Write in English, since that's the language of the training data described in the model card.

Can you use PAN 2 for work?

Yes: PAN 2's code and weights are released under the Apache 2.0 license, which allows commercial use in Italy and the European Union too. That's an important difference from other recent models, such as MiniMax H3-Image, whose license excludes the EU. Just pay attention to the third-party components (the Wan 2.2 VAE and Qwen3.5-9B), which have their own licenses: the official model card points to the NOTICE and THIRD_PARTY_LICENSES files. Our guide to AI video models in ComfyUI compares it with Wan, LTX and the others, licenses included.

Try it right away

If you have a card with 24 GB or more and at least 64 GB of RAM, this is the shortest route:

  1. update ComfyUI to v0.46.0 or later;
  2. download from Comfy-Org/PAN2 the int8 model, the int8 Qwen3.5-9B encoder and the Wan 2.2 VAE, each into its folder;
  3. run a short, small first test (few frames, reduced resolution) to measure the times on your card;
  4. only then move to 241 frames, with a structured prompt like the one above.

To see results before downloading over 37 GB of files, watch two videos released in the last few days: "PAN2 A Local World Video Model Know About Physics! Try It In ComfyUI" by Benji's AI Playground and "PAN 2: NEW AI World Model! Full Tutorial in ComfyUI (T2V & I2V)" by The Ai Blueprint. With 12 GB of VRAM, models like LTX or Wan 2.2 in GGUF form remain more practical for local video; for talking characters there's LongCat Avatar 1.6, the talking avatar from a photo and voice.

Other news

  • Suno Visuals: since November 17 Suno creates video clips up to 15 seconds from your songs, plus lyric videos and cover art. Video credits are separate from music credits (Suno's announcement).
  • MAI-Image-2.7: at Ignite on November 17 Microsoft introduced seven new MAI models, including this one for creating and editing images, with a cheaper Flash version (Thurrott).
  • ComfyUI v0.46.0: besides PAN 2, the November 18 release adds the Split Image Channels node to separate an image's channels and brings Qwen3-1.7B to the Generate Text node (official changelog).

Frequently asked questions

Is PAN 2 free?

Yes. Code and weights can be downloaded from Hugging Face under the Apache 2.0 license, and the model runs on your computer, with no subscriptions or credits. The real cost is hardware: you need a graphics card with lots of memory, plenty of RAM and at least 37 GB of free disk space for the int8 version.

Does PAN 2 run on a 12 GB card?

There's no official figure and we haven't tried it. The int8 model weighs 26.7 GB, more than double the memory of an RTX 3060, and has to be loaded in pieces from RAM together with the 9.75 GB text encoder. With 32 GB of RAM it's at the limit; until a lighter GGUF version comes out, we don't recommend it.

How long is a PAN 2 video?

Up to 20 seconds in a single clip, meaning 481 frames at 24 fps. The default is 241 frames, about 10 seconds. The standard resolution is 704×1248. On an RTX 5090 a 10-second clip takes about 12-13 minutes of sampling with the recommended 50 steps.

Do you need a custom node for PAN 2?

No. Since the PR merged on November 15, PAN 2 works with standard ComfyUI nodes: Load Diffusion Model, Load CLIP with the pan2 type, KSampler and VAE Decode. The community ComfyUI-PAN2 pack is still handy for its Prompt Builder, which helps you write the structured prompt the model needs.

If you want to follow the tests we actually run on our RTX 3060, subscribe to the @tutorial3d YouTube channel and grab the free workflows from the ComfyUI section of the site.

Tutorial3d
Grafico - Videodesigner - VisualFX

Ti consigliamo anche:



#Devo averlo!

VIDEO

INSPIRATION IN YOUR INBOX

Una mail. Un tutorial. Niente rumore.

Ti scrivo un solo tutorial scelto, con file progetto e note tecniche. Senza promozioni, senza affiliate. Cancellabile in un click.

[newsletter_form]