Wan 2.2 is a video generation model family, and this workspace lets you use it without touching a terminal. You type a description, choose a few options, and the render happens on our hardware rather than yours. Nothing is installed on your machine, and you do not need a discrete graphics card to get a result. If you want the broader picture of what the family covers, the Wan overview is the place to start, and the Wan 2.2 hub collects the generation-specific pages.
The 2.2 generation is a text-and-image conditioned video model. In practice that means two common entry points: describing a scene in words, or handing over a still image and letting the model move it. Both routes live here as separate tools, so you can pick text to video when you are starting from nothing but a sentence, or image to video when you already have a frame you like. There is also a route for driving motion from a reference clip, which you can read about under reference to video.
What this site is not: a download mirror. We do not distribute model weights, checkpoints, quantised files or safetensors, and there is nothing here to install locally. If your goal is to run the model on your own machine, pages like Wan 2.2 GGUF or Wan 2.2 Hugging Face describe that world, but the actual files come from elsewhere. Here you get a browser form and a rendered clip.
The single biggest factor is the prompt. Vague prompts produce vague motion. Name the subject, what it is doing, where it is, and how the camera behaves. "A woman walks" gives the model almost nothing; "a woman in a red coat walks toward the camera along a wet pavement, shallow depth of field, slow forward dolly" gives it a plan. Our Wan 2.2 prompt notes go deeper on structure if you want a repeatable formula.
Treat the first render as a test, not a deliverable. Watch it once to see whether the composition and motion direction match your intent, then adjust one thing at a time. Changing the prompt, the seed and the length all at once makes it impossible to learn what actually helped. Short clips are cheap to iterate on and tell you more per minute than a long render that misses the point.
If you are starting from a still, the source image matters as much as the text. A clean, well-lit frame with a clear subject gives the model far less to invent. You can tidy an input with the image compressor or convert formats with the image converter before uploading.
Aspect ratio and resolution decide the frame you are working in, and they are hard to fix after the fact. Pick the shape that matches where the clip will live rather than cropping later. The resolution page covers how the 2.2 generation handles different output sizes.
Duration is the next lever. Longer clips give the model more room to drift, so a tight four-second shot usually holds together better than a sprawling one. You can always stitch or extend afterwards; the video extend tool exists for exactly that.
The seed controls repeatability. Fix it when you are refining a composition and want to see only the effect of a prompt change. Let it vary when you are exploring and want to see the range of interpretations.
The cost of a render is displayed alongside the form before you commit, so you can see what a given configuration will use up rather than guessing. Plans and what they include are laid out on the pricing page.
This is not a video editor. There is no timeline, no multi-track audio, no colour grading suite. If you need to cut between shots or sync dialogue, render the pieces here and assemble them in an editor.
It is not a source of model files. Nothing on this site is downloadable as a checkpoint, and no local inference is possible through it. If you need weights, look to the model's own distribution channels.
It is also a poor fit for precise, literal reproduction of text or logos inside a frame. Generative video models in this class are not built for typographic accuracy, and asking for it usually wastes renders. Use a still image tool for that kind of work and animate the result instead.
Finally, do not expect a first pass to be broadcast-ready. These are drafts. The workflow that works is generate, review, refine, and only then commit to a longer or higher-cost run.
Do I need a GPU to use this?
No. The rendering happens on our infrastructure. Your browser sends the prompt and options, and you get a clip back. A modest laptop is enough to drive the interface.
Can I download the Wan 2.2 model here?
No. We do not host or distribute weights, checkpoints or quantised files. This is a hosted generation workspace, not a model repository.
How long should my first clip be?
Start shorter than you think you need. A brief clip renders faster and tells you whether the prompt is working before you spend more on a longer run.
What if the motion looks wrong?
Change one variable at a time, usually the prompt first. Describe the camera movement and the subject's action explicitly, then re-render and compare against the previous attempt.