ComfyUI: Make Free AI Videos on Your Own Laptop — Wan 2.2 Guide (No Subscription)

Every AI video tool on the web charges you per clip — credits, subscriptions, watermarks. ComfyUI is the way out: a free, open-source app that runs AI video generation on your own computer, with no subscription, no credits and no usage limits. Pair it with Alibaba's open-source Wan2.2 video model and you can turn a text prompt — or a single photo — into a 720p video clip, on your own machine, as many times as you like. This guide shows you exactly how, including the honest hardware requirements and the one trap to avoid.

What ComfyUI is

• Free forever, open-source (GPL-3.0): the desktop app costs nothing, needs no account, and runs entirely on your machine — offline after setup. Those are the words on the official download page, checked today.
• Node-based, no code: you build a workflow from small blocks — load a model, type a prompt, generate, save — by connecting them on screen. It opens in your browser at http://127.0.0.1:8188 once it is running.
• Not a website: there is nothing to sign up for and no queue. Your GPU does the work, so your clips are private and unlimited.
• The video model in this guide is Wan2.2-TI2V-5B — text or image to video, about 5 seconds per clip (121 frames at 24 fps), 720p, Apache 2.0 licence, from Alibaba's Wan team. Its official page is on Hugging Face at huggingface.co/Wan-AI/Wan2.2-TI2V-5B.

Desktop vs Cloud — do not mix these up

ComfyUI has two faces and only one of them is free. Comfy Desktop — the app you download — is free forever. Comfy Cloud is a separate hosted service that runs on credits and is paid. If a page asks you to buy credits, you are on the cloud product, not the desktop app. This guide uses the desktop app only.

What you need — the honest version

Video generation is heavy work, and this is the part other guides skip. The Wan2.2 README is clear:

• Best case: a GPU with 24 GB of VRAM (for example an RTX 4090) runs the TI2V-5B model at full 720p quality.
• Workable: around 8 GB of VRAM with a quantised build — slower, but it works.
• Too slow to enjoy: a 6 GB card. In one community test, an RTX 3050 with 6 GB took about 5 minutes to make a 1-second clip.
• Disk space: keep 30 GB or more free. Video models are large files, 5–40 GB each.
• The app itself needs Python and PyTorch 2.4.0 or newer, but the desktop installer handles that for you — you do not install them by hand.

Step 1: Install ComfyUI Desktop

1. Go to comfy.org/download and download the installer for your system — Windows (x64 or ARM64), macOS (Apple Silicon) or Linux (AppImage). The page detects your platform automatically.
2. Run the installer and open Comfy Desktop. It starts a local server and opens its interface in your browser. Nothing is uploaded anywhere.
3. If you prefer the classic route, the portable build is on GitHub at github.com/comfyanonymous/ComfyUI — clone it and install the requirements from the README.

Step 2: Load the Wan 2.2 video workflow

1. Download the TI2V-5B model weights from the Wan2.2-TI2V-5B Hugging Face page and place them in ComfyUI's models folder, as the model page instructs. No Hugging Face token is needed — the repo is public.
2. In ComfyUI, load a first-party Wan 2.2 workflow template from comfy.org/workflows. A template is a ready-made chain of nodes, so you do not have to build the workflow yourself.
3. Check the model loader node points at the TI2V-5B weights you downloaded. If your GPU is on the smaller side, enable the memory-saving options in the workflow (model offload and the low-VRAM settings) — the Wan2.2 README uses --offload_model True and --t5_cpu for the same reason.

Step 3: Make your first video

1. For text-to-video, type a plain, specific prompt — subject, action, setting, light, camera. Example: "a white cat wearing sunglasses sits on a surfboard, close-up, bright seaside light, gentle camera push-in".
2. For image-to-video, attach a photo first and describe the motion you want in the prompt. The image sets the scene; the prompt sets the movement.
3. Press Queue Prompt and wait. A 5-second 720p clip on a 24 GB card takes minutes, not seconds — video is the slowest kind of AI generation. Save the result from the output node.
4. Start with the 480p setting for your first test, then step up to 720p once the workflow runs cleanly. Short prompts with one clear action give cleaner motion than long, complicated ones.

No strong GPU? Read this before you click anything

Honesty first: as of today, 14 November 2026, there is no working free in-browser demo of Wan2.2 to send you to. The official Hugging Face demo Space (Wan-AI/Wan-2.2-5B) no longer exists — the page returns a 404. The community page at huggingface.co/spaces/wavespeed/wan2.2 is only a model reference page with no generator on it; its "Run it" section points to WaveSpeed's API, which needs an API key and is a paid service. Do not buy API credits thinking that is the free demo. The genuinely free route is the local one above — ComfyUI on your own GPU. If your machine cannot run it, the practical option is a borrowed or cloud machine with a 24 GB GPU, not a "free" web demo that quietly charges you. We will update this section if a real free demo appears.

What it cannot do

This is not a one-click video app. The first setup — app install plus a multi-gigabyte model download — takes real time, and you need to learn how a workflow fits together before it feels easy. Clips are about 5 seconds each; longer videos mean chaining clips together in an editor. And the bigger 14B Wan2.2 models (T2V-A14B, I2V-A14B) need an 80 GB GPU for a single-card run — far beyond a home machine — which is exactly why this guide uses the 5B model. If you want a quicker start on a smaller machine, the community LightX2V project distils Wan2.2 for faster generation, but that is an optimisation for local runs, not a separate free service.

Bottom line

ComfyUI plus Wan2.2 is the most capable free AI video setup we have covered: no subscription, no credits, no watermark, no queue — your laptop, your GPU, your clips. The price is honest hardware requirements and an afternoon of setup, not money. If you have a gaming GPU with 8 GB or more, this is the weekend project that replaces a paid video subscription. If you do not, do not pay for an API and call it free — wait for a real demo, or borrow the GPU.

Sources: comfy.org/download (checked live 14 November 2026), the Wan2.2 GitHub README and the Wan2.2-TI2V-5B model page. The Hugging Face demo Space status was checked live on 14 November 2026. Hero photo: a film camera at Cinecittà Studios, Rome, by Sonse, CC BY 2.0, via Wikimedia Commons.