DEV Community

Wan2.2: Free Open-Source AI Video Generation - 720p Cinematic Clips With No Subscription

Most AI video tools charge you per clip. Wan2.2 does not. It is Alibaba's open-source video generation model - Apache 2.0 licensed, free to use including commercially, with more than 18,800 stars on GitHub - and its light 5B model makes real 720p video on a normal consumer graphics card.

This is the complete guide: what it does, what it needs, and how to run it free, with or without a strong PC.

What is Wan2.2?

Wan2.2 is the second generation of Alibaba's Wan video models, released in July 2025 with its inference code and model weights fully open. It comes in three main versions:

  • TI2V-5B - one 5-billion-parameter model that does both text-to-video and image-to-video at 720p and 24 frames per second. This is the one most people should start with: the official README says it can generate a 5-second 720p video in under 9 minutes on a single consumer-grade GPU.
  • T2V-A14B - a 14-billion-parameter text-to-video model, and I2V-A14B for image-to-video, both at 480p and 720p. These use a Mixture-of-Experts design - two expert models, one for the noisy early steps and one for the fine details - for higher quality, but they need far more GPU memory.
  • S2V-14B and Animate-14B - bonus models in the same family: S2V turns speech or audio into a talking video, and Animate moves a character in a photo using a template video.

The licence matters: Apache 2.0 means you can use the model freely, modify it, and use what you generate commercially, as long as you keep the licence notice. The model makers state they claim no rights over your generated content.

What you need

For the recommended route - the 5B model in ComfyUI - you need a Windows, Mac or Linux PC with an NVIDIA graphics card, plus Python and about 20 GB of free disk space for the model files.

The 5B model is the light one; the 14B models are a different league and need a serious GPU, so skip them unless you know your card can handle them. If you have no GPU at all, use the free Colab route below instead.

Route 1: Run it on your own PC with ComfyUI (free, full control)

ComfyUI is the free, open-source node-based tool for running image and video models locally, and Wan2.2 is officially integrated into it. This is the route the ComfyUI team documents step by step.

  • Install ComfyUI from github.com/comfyanonymous/ComfyUI and launch it - it opens in your browser at a local address.
  • Download the three model files from the official repackaged set at huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged: the diffusion model wan2.2_ti2v_5B_fp16.safetensors, the text encoder umt5_xxl_fp8_e4m3fn_scaled.safetensors, and the VAE wan2.2_vae.safetensors. Put them in ComfyUI's models/diffusion_models, models/clip and models/vae folders.
  • In ComfyUI, load the built-in Wan2.2 5B workflow template (Workflows > Browse Templates > Video > Wan2.2 5B). Check the three Load nodes point at the files you downloaded.
  • Write your prompt in the CLIP Text Encoder node - one clear action, calm pacing, e.g. "A golden retriever puppy runs through autumn leaves in a sunny park, cinematic lighting, smooth camera follow." Keep the default size and frame count for your first try.
  • Press Queue Prompt (Ctrl+Enter) and wait. A 5-second 720p clip takes a few minutes on a consumer GPU - the README's figure is under 9 minutes. Your video appears in ComfyUI's output folder.
  • For image-to-video instead: enable the Load Image node (Ctrl+B), upload a still image, and the model animates it. A strong still gives you a much better result than text alone.

Route 2: No GPU? Run it free on Google Colab (T4 GPU)

If your PC has no NVIDIA card, Google Colab's free tier gives you a Tesla T4 GPU in the browser - no install, no bill. Free Colab GPUs are not guaranteed and sessions time out when idle, so save your clips to Google Drive as you go.

  • Open colab.research.google.com, start a new notebook, then go to Runtime > Change runtime type and pick T4 GPU. Run !nvidia-smi in a cell to confirm the GPU is there.
  • Install ComfyUI in the notebook and the ComfyUI-GGUF custom node - then load the community GGUF quantised Wan2.2 models, which are built specifically for the free T4: huggingface.co/geceff/Wan2.2-Custom-Models-GGUF. Quantised means the model is compressed so it fits in the T4's 15 GB of memory; the makers warn not to force full-size models through the Colab GUI or it will run out of memory.
  • Follow the same prompt and generate steps as Route 1, with the smaller default size and short clip length the GGUF guide recommends, then download your video from the notebook.

Honest limits - read this before you start

  • The official browser demo is currently paused. The Wan team's own Hugging Face Space for the 5B model is paused, so there is no official one-click web demo right now. Community Spaces exist, but they come and go and queue during busy hours - the two routes above are the reliable free paths.
  • Prompt extension costs extra (or use a local model). Wan2.2 can expand your short prompt into a detailed one automatically, but the easy method needs a DashScope API key. You can skip it - short, well-written prompts work fine - or use the local Qwen model option in the README instead.
  • The 14B models are not for weak hardware. The README's own example command for the 14B text-to-video model expects a GPU with at least 80 GB of VRAM unless you use its offload options. Start with the 5B model; it is the one built for ordinary computers.
  • Generation is slow, and clips are short. Five seconds at 720p in a few minutes is normal for free local generation. Plan for short clips and stitch them together, rather than expecting a minute-long video in one go.

Bottom line

Wan2.2 is the strongest free, open-source video model most people can actually run today: a real 720p text-to-video and image-to-video model, Apache 2.0 licensed, with an official ComfyUI workflow and a free Colab path for anyone without a GPU. Start with the 5B model, keep your first prompts simple, and try one image-to-video clip from a photo you already like - that is the fastest way to see what it can do.

Links: Wan2.2 on GitHub ยท ComfyUI Wan2.2 official tutorial ยท Wan2.2-TI2V-5B on Hugging Face ยท T4-optimised GGUF models for Colab.

Hero photo: film set in Moscow, by Lesless, CC BY-SA 4.0, via Wikimedia Commons.

Originally published on Ai Game Changer Pro: https://aigamechangerpro.blogspot.com/2026/11/wan22-free-open-source-ai-video.html

Read on DEV Community ↗ ← Back to News

Comments

No comments yet. Start the discussion.