AI video generators are no longer just cloud services. Open models can now turn a text prompt, or a still image, into a short clip on an ordinary gaming PC, with no monthly fee and nothing uploaded. The most accessible starting point is Wan 2.2, an open-source video model from the Wan project, running in the free ComfyUI app.
The steps below follow ComfyUI's official Wan 2.2 tutorial and the Wan 2.2 project page on GitHub. Expect generation to take minutes rather than seconds, and to vary a lot with your graphics card.
What you need
A graphics card with at least 8GB of VRAM. ComfyUI's tutorial says the Wan 2.2 5B model "should fit well on 8GB vram with the ComfyUI native offloading". More VRAM means less offloading and faster results. Our VRAM guide explains the memory tiers.
Plenty of system RAM, because offloading moves parts of the model into normal memory. Our RAM for local AI guide covers how much to aim for.
ComfyUI, kept up to date. Install it with our ComfyUI on Windows guide, or with Stability Matrix if you want a one-click installer. ComfyUI warns that the Wan 2.2 templates may be missing if your copy is out of date.
Some free disk space for the model files.
Which Wan 2.2 model to use
ComfyUI's documentation lists three main Wan 2.2 models:
Model | What it does | Who it suits |
|---|
Wan2.2-TI2V-5B | Text-to-video and image-to-video in one model | Most home PCs; the one to start with |
Wan2.2-T2V-A14B | Text-to-video | High-end cards |
Wan2.2-I2V-A14B | Image-to-video | High-end cards |
The Wan project says the 5B model supports 720p video at 24 frames per second and "can also run on consumer-grade graphics cards like 4090". The larger A14B models use a mixture-of-experts design. The Wan team's own example command for running them on a single GPU says it "can run on a GPU with at least 80GB VRAM", although ComfyUI's page also points to GGUF versions, which are generally smaller. Start with the 5B model.
Wan 2.2 is released under the Apache 2.0 licence, according to its GitHub page.
Step 1: Load the Wan 2.2 5B template
Open ComfyUI and go to the Workflow Templates (template library).
Search for Wan2.2 5B and open the "Wan 2.2 5B Video Generation" template. ComfyUI's tutorial also offers the workflow as a JSON file to download.
If ComfyUI reports missing nodes, update it and try again.
Step 2: Download the three model files
ComfyUI's tutorial lists three files for the 5B workflow and where each one goes inside your ComfyUI folder:
File | Folder |
|---|
wan2.2_ti2v_5B_fp16.safetensors | ComfyUI/models/diffusion_models/ |
wan2.2_vae.safetensors | ComfyUI/models/vae/ |
umt5_xxl_fp8_e4m3fn_scaled.safetensors | ComfyUI/models/text_encoders/ |
The template normally prompts you to download them. If not, the tutorial links each file. Restart ComfyUI or refresh the model lists once they are in place.
Step 3: Write a prompt and generate
Check that the Load Diffusion Model, Load CLIP and Load VAE nodes show the three files above.
Type your description into the prompt box. Be specific about the subject, the camera movement and the lighting.
In the video size node you can change the resolution and the length, which is the total number of frames. ComfyUI's tutorial points to these settings; shorter and smaller clips are much quicker.
Click Run and wait.
Image-to-video
The same 5B model can animate a still picture. ComfyUI's tutorial says you can press Ctrl+B to enable the Load image node, then upload the image you want as the first frame.
Tips if it is slow or runs out of memory
Start small. Generate a short, lower-resolution clip first, then scale up once you like the result.
Close other GPU-hungry apps, including games and browsers with many tabs. Our guide to checking VRAM usage on Windows shows what is using your graphics memory.
Try a GGUF or FP8 version of the larger models if you have a high-end card. ComfyUI's tutorial lists GGUF versions.
Smooth the result. NVIDIA said in October that new RTX Video nodes for ComfyUI can turn 15 frames-per-second AI video into smoother 60fps footage on its hardware. We have not tried them.
ComfyUI or an easier app?
ComfyUI's node graph takes some getting used to. If you mainly want still images, our ComfyUI vs Amuse comparison explains the simpler alternatives, and our Stable Diffusion and Flux PC guide covers image hardware.
Our verdict
The easiest way to try AI video at home is the open Wan 2.2 5B model in ComfyUI: load the "Wan 2.2 5B" template, put the three model files in the folders ComfyUI lists, write a prompt and run it. ComfyUI says the 5B model fits on an 8GB graphics card using its built-in offloading, while the larger 14B models need high-end hardware. Keep clips short and small at first.