Wan 2.2 video to video means handing the model a clip you already have instead of a blank prompt, then letting it rework, restyle, or continue that footage. On wan2-6.net this happens through two routes: reference, where your clip guides a fresh generation, and extend, where the model picks up where your footage stops. Both run in the browser on hosted GPUs, so nothing is installed on your machine and no local hardware is required.
If you are still deciding which Wan generation fits your project, the Wan overview lays out the family, and Wan 2.2 covers this version specifically.
The reference route treats your uploaded clip as visual guidance. The model reads its motion, framing, and colour, then produces new frames that follow that lead rather than starting from nothing. You are not editing the original file frame by frame; you are generating a companion piece shaped by it.
The extend route is narrower and more literal. It takes the tail of your clip and continues forward, aiming for a seamless join in motion and tone. Use it when you want more seconds of the same shot, not a reinterpretation.
Neither route is a download. This site does not distribute model weights, checkpoints, or installers, and there is nothing here to run offline. If you came looking for a local setup, the hosted form is the only path offered.
Start with clean source footage. Compression artefacts, heavy grain, and shaky handheld motion all get inherited and amplified, because the model is reading your clip as a specification. A short, stable, well-lit segment beats a long messy one every time.
Keep your first attempt modest. Treat the opening render as a test rather than a finished asset: check whether the motion reads correctly, whether the colour survived, and whether the join in an extend lines up. Only after that should you spend time on a longer or higher-quality pass.
Write your prompt as a description of what should change, not a restatement of what is already there. If the source is a person walking through a market, the prompt does not need to say "person walking through market" — it needs to say what the new version should look like instead.
For phrasing ideas, Wan 2.2 prompts collects patterns that work well with this model family.
The reference route and the extend route are not interchangeable, and picking the wrong one wastes a run. Reference is for transformation; extend is for continuation. If your goal is a longer version of the same shot, extend. If your goal is the same shot reimagined, reference.
Source length matters more than source resolution in most cases. A tight, coherent few seconds gives the model a clearer signal than a sprawling minute.
Prompt strength is the other lever. Push it too hard and the model abandons your source footage entirely, producing something that ignores the input. Push it too soft and you get a near-copy with no visible change. The useful range is usually narrower than people expect.
Before you commit, the interface displays what the run will cost in credits next to the submission form, so you can see the figure for your chosen settings before anything is charged. Check it there rather than guessing.
Video to video is a poor fit for precise editing. If you need to remove an object, swap a background cleanly, or match a specific frame exactly, this is the wrong tool — the model generates, it does not retouch. For still-image work of that kind, image edit is the appropriate page.
It also struggles with very long clips. The longer the input, the more the model drifts, and extend chains compound that drift with each pass. If you need minutes rather than seconds, plan on stitching segments and accepting some inconsistency at the seams.
Finally, it will not rescue bad footage. A clip that is out of focus, badly exposed, or full of motion blur will produce output with the same problems. Fix the source first, or choose different footage.
Can I use video to video without uploading anything?
No. Both the reference and extend routes need source footage to work from. If you want to start from text alone, text to video is the right entry point.
Does it keep the audio from my clip?
The generation focuses on the visual track. Plan on handling sound separately, and check the output before assuming any audio carried through.
How many attempts should I expect?
More than one. The first pass tells you whether your source and prompt are pulling in the same direction, and most people adjust after seeing it.
Is there a way to run this locally instead?
Not through this site. We host the models and serve the results; we do not provide weights, downloads, or offline packages. If local execution is a hard requirement, this is not the service for it.